Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Munich Personal RePEc Archive
Networks, Frictions, and Price Dispersion
Donna, Javier and Schenone, Pablo and Veramendi, Gregory
The Ohio State University, Arizona State University, Arizona State
University
17 July 2015
Online at https://mpra.ub.uni-muenchen.de/66999/
MPRA Paper No. 66999, posted 02 Oct 2015 10:29 UTC
Networks, Frictions, and Price Dispersion
Javier Donna Pablo Schenone Gregory Veramendi
The Ohio State University Arizona State University Arizona State University
First version: January, 2014
This version: July 17, 2015
We are grateful to Rune Vejlin for the initial conversations which led to this project. Discussions withMatt Backus, David Blau, Hector Chade, Eleanor Dillon, Laura Doval, Domenico Ferraro, John Hatfield,John Kagel, Shachar Kariv, Rasmus Lentz, Dan Levin, Jim Peck, Tiago Pires, Rob Porter, Andrew Rhodes,Huanxing Yang as well as seminar participants at Arizona State University, Cycles, Adjustment, and Policyconference (Aarhus University), the Southwest Search and Matching conference (University of CaliforniaRiverside), and the 13th International Industrial Organization conference (Boston) have greatly benefitedthis work. All errors are unintentional.
Abstract
This paper models frictions in buyer-seller markets using networks, where buyers are
linked with a subset of sellers and sellers are linked with a subset of buyers. Sparser
networks are associated with higher search frictions. We use the model to characterize
pairwise stable allocations and their supporting prices. Our approach allows for network
effects, where a buyer who is not linked to a seller affects the price obtained by that seller.
Network effects generate the central finding of our paper: even relatively sparse networks
lead to price distributions and allocations that are close to the perfectly competitive
outcome where the law of one price holds. We then investigate the role of network
effects in a dynamic setting by studying wages in the context of an on-the-job search
model. We find two novel predictions relative to the search literature. Lowering frictions
(so that workers receive job offers at a higher rate) leads to: (1) lower worker mobility
and lower expected wage growth and (2) lower expected wages in markets with high
unemployment. We argue that our framework is suited to the analysis of a wide range
of real-world markets, such as the labor market and buyer-seller trading platforms like
eBay or Amazon.
JEL Codes: C78; D44; L00; J31
Keywords: Networks; Matching; Auctions; Competition; Frictions; Price Dispersion
Javier Donna Pablo Schenone
Department of Economics Department of Economics
The Ohio State University W.P. Carey School of Business
1945 N High St Arizona State University
417 Arps Hall 501 E. Orange Street
Columbus, OH 43210 Tempe, AZ 85287
Phone: 614-688-0364 Phone: 480-965-5596
Email: [email protected] Email: [email protected]
Gregory Veramendi
Department of Economics
W.P. Carey School of Business
Arizona State University
501 E. Orange Street
Tempe, AZ 85287
Phone: 480-965-0894
Email: [email protected]
1 Introduction
Price dispersion is observed in many markets even after accounting for observable charac-
teristics. Examples include labor markets, where similar workers are paid different wages
(Mortensen 2005); buyer-seller trading platforms such as eBay or Amazon, where identical
goods are sold by the same seller at different prices (Einav, Kuchler, Levin, and Sundaresan
2015); and markets for automobiles, where identical automobiles are sold at different prices
by the same dealer (Morton, Zettelmeyer, and Silva-Risso 2001). A common characteristic
of these markets is that buyers interact with multiple, but not all, sellers. In the case of
labor markets, firms interview multiple applicants. Bidders on eBay bid in multiple auctions
for an identical product. Consumers visit multiple automobile dealerships before making a
purchase. A natural question arises: How does the fact that buyers interact with multiple
sellers affect price dispersion?
The central insight of this paper is that “indirect competition” plays an important role
in determining price dispersion. For example, consider the case of two eBay auctions where
there is only one common bidder participating in both auctions. The bidders in one auction
indirectly compete with the bidders in the other auction because the two auctions are con-
nected through the common bidder.1 Allowing for indirect competition results in what we
call “network effects”: an interdependence in the prices between these two auctions caused by
indirect competition. Even if many sellers are not directly competing for the same buyer, net-
works effects can equalize prices across them. How buyers and sellers are connected (linked)
determines the extent of the network effects and motivates the use of networks for our analy-
sis. We study prices and allocations in a network using pairwise stability as our matchmaking
criterion since this is the weakest criterion for matchmaking that is consistent with Pareto
efficiency. To the best of our knowledge, there has been no attempt to study price disper-
sion in networks using pairwise stability as the matchmaking criterion.2 See Section 2 for a
detailed literature review.
In this paper, we use networks to model buyer-seller markets where a buyer can obtain a
good from the seller only if the two are linked. Perfect competition assumes that all buyers
are linked to all sellers in the network, leading to the Walrasian outcome. Frictions are present
in the market whenever there is at least one seller that is not linked to every buyer. Hence,
the level of frictions in the network is determined by the total number of links or sparsity
of the network. In this setting, we characterize pairwise stable allocations restricted to the
network and the set of prices that sustain them. While we characterize allocations and prices,
computing these for a large network is not tractable. Thus, we develop a computationally
tractable algorithm that finds the upper and lower bounds of the set of prices that sustain
any pairwise stable match.
1See section 3 for an example with buyers and sellers.2The only other paper that we are aware that uses pairwise stability as the matchmaking criterion in a
network with frictions is Elliott (2015).
1
A central finding of our paper is that even relatively sparse networks lead to price distri-
butions and allocations that are close to the perfectly competitive outcome.3 For example, in
a network with 10,000 sellers, over 99% of the sellers are paid the same price when less than
0.1% of the possible links are active. The prediction that even sparse networks lead to price
profiles that are close to the perfectly competitive outcome is a consequence of network ef-
fects. As the number of links increases, indirect competition among buyers rapidly becomes
more likely. This indirect competition pushes towards price equalization in markets with
a homogeneous good. Indirect competition makes the market look “as if” it was perfectly
competitive, hence the result.
We then investigate the role of indirect competition in a dynamic setting by studying
wages in the context of an on-the-job search model.4 In a standard on-the-job search model,
employed workers can search for better job offers, while firms cannot search to replace a
worker with another one at lower wage. This creates an asymmetry where employed workers
always benefit from lower frictions through higher wage growth (workers receive offers at a
higher rate) and higher wages in the overall economy (Pissarides, 1990). An implication is
that new search technologies (e.g. the internet) promise wage gains to workers. In addition,
labor market policies designed to reduce frictions (e.g. job search assistance programs5) are
intended to not only help workers find jobs faster, but also help them attain higher wages.
Our model predicts that workers do not always benefit from reducing frictions due to network
effects. The benefit varies depending on whether the markets are tight (fewer workers than
open vacancies at firms) or loose (more workers than vacancies). In loose markets, lowering
frictions improves the outside options of firms. Although we include the same asymmetry as
in on-the-job search models, firms receive multiple links and are more likely to have a second
link with an unemployed worker. This drives expected wages down. In turn, employed
workers are more likely to compete with unemployed workers, limiting wage growth and job-
to-job mobility. In tight markets, workers benefit from reduced frictions as in a standard
search model. Firms compete more strongly for workers, driving wages up. Workers have
better outside options and receive a higher initial wage out of unemployment, but this leaves
little room for wage growth. Likewise, since workers start off with a higher wage, they
have a lower probability of moving to a job that pays more. Hence, even in tight markets,
lowering frictions diminishes wage growth and job-to-job mobility. Although the asymmetry
3Our finding that pricing behavior in networks with frictions closely resembles the perfectly compet-itive outcome is consistent with the pricing behavior observed in the laboratory experiments in Charness,Corominas-Bosch, and Fréchette (2007) and Gale and Kariv (2009). See Judd and Kearns (2008) for a surveyof experiments in networked markets.
4Search models can be mapped into the corresponding network using the same firms and workers whereeach worker receives a link from a firm when they receive a job offer from that firm in the search model.Many search models are in continuous time. Hence, decreasing frictions in these models increases the offerarrival rates to workers. Yet, since these models are in continuous time, at any instant they only receive oneoffer.
5Gautier, Muller, van der Klaauw, Rosholm, and Svarer (2014) estimate the equilibrium effects of a jobsearch assistance program in Denmark.
2
between workers and firms is still present in our model, network effects mitigate its impact
in loose markets. Hence, our model makes new predictions about the effect of technologies
and policies that affect the level of frictions in the economy.
Our paper contributes to the literature on networks. The networks literature typically
proposes a concrete buyer-seller game to be played within these networks and identifies
conditions under which the equilibria of these games are Pareto efficient.6 In contrast, we
use pairwise stability as the matchmaking criterion and focus on how the distribution of prices
depends on the level of frictions in the market. Using pairwise stability as our matchmaking
criterion implies that the equilibrium of any game that is consistent with Pareto efficiency
will be included in our set of matches.
In summary, we make four main contributions: (1) we characterize the set of prices that
sustain any pairwise stable matching in an unrestricted network; (2) we develop an algorithm
for finding a stable matching and its supporting prices that is tractable for large networks;
(3) we study price dispersion in arbitrary sparse networks; and (4) we study wage dispersion
and wage growth in an application to labor markets.
Outline of the Paper
The theoretical analysis begins by studying arbitrary buyer-seller networks. These networks
are exogenously formed by linking buyers and sellers, but no restrictions are placed on how
many sellers a buyer can be linked with. Sellers offer one unit of an indivisible good. Buyers
have single unit demand and also differ in their valuation of the good. To keep the model
simple, we focus on the homogeneous goods case, where all sellers have the same valuation
for the good. Buyers’ utility is their valuation less the price if they obtain the good and zero
otherwise. Sellers’ utility is the price they are payed if they sell the good and their valuation
otherwise.
In the model, we assume matches form according to pairwise stability restricted to the
network. That is, a buyer obtains a good from a seller at a price p if four conditions hold:
first, the buyer and the seller are connected in the network; second, there is no other seller
linked to this buyer that is willing to sell at a price lower than p; third, there is no other
buyer linked to the seller that is willing to pay a price higher than p; finally, the price p lies
between the seller’s valuation and the buyer’s valuation. This is the weakest criterion for
matchmaking that is consistent with Pareto efficiency.
The main theorem characterizes the set of all pairwise stable matchings in an arbitrary
network and the set of prices that sustain them. To do that we decompose the original network
into a network of fully connected subnetworks. With this, we identify two components
that jointly determine the prices that sustain pairwise stable matchings: a pure competitive
6See for example Kranton and Minehart (2001), Corominas-Bosch (2004), Manea (2011), Gautier andHolzner (2013), and Elliott (2014). Elliott (2015) uses pairwise stability as the matchmaking criterion in anetwork with frictions, but does not investigate price dispersion.
3
component and an outside option component. Fully connected subnetworks are competitive
markets with a unique price. The links between these subnetworks introduce an outside
option component since a buyer from one subnetwork might choose to participate in another.
The final prices that sustain a stable matching can be thought of as the result of these
two effects. The theorem is useful because it allows us to rationalize any pairwise stable
matching as an outcome of a decentralized market. In particular, we show that the outcome
of a mechanism that generates any pairwise stable matching is equivalent to the outcome
of a set of independent second-price auctions. We show how one can construct a set of
independent auctions for any pairwise stable matching.
We use our characterization of the prices that support pairwise stable matchings to design
an algorithm which we use to simulate large markets. For any given pairwise stable matching,
calculating the set of prices that sustain it is simple whenever the network is small (e.g. 2
sellers and 3 buyers). However, since the applications of interest (the labor market, inter-
net auctions) involve large economies, calculating the set of sustaining prices is intractable.
Hence, we design an algorithm that outputs a matching (i.e a complete specification of
buyer-seller matches) that is pairwise stable, and the prices that sustain it.
The algorithm we use to simulate the model is a deferred acceptance algorithm that works
in two stages. In the first stage, one side of the market (e.g. sellers) hold “auctions” and the
other side (e.g. buyers) sequentially “bid” in their linked auctions. Once no bidder wants to
make any new bids, the algorithm ends. A corollary of the main theorem is that the outcome
of the algorithm is indeed a stable matching. The second stage of the algorithm calculates two
sets of prices: (1) the minimum prices that just price out the unmatched agents and support
the stable matching; and (2) the maximum prices that give all matched agents nonnegative
utility and also support the stable matching. In this way, the second stage finds the lower
and upper bounds of the set of prices that support the pairwise-stable matching from the
first stage.
We simulate a range of networks to obtain predictions about price distributions. We start
with a set of heterogeneous buyers and homogeneous sellers. We parameterize the level of
frictions by choosing the number of links per buyer in the network. This determines the
total number of links in the network. The simulation then randomly draws links between
sellers and buyers. After the network is realized, we run the algorithm and generate a price
profile. Given that our algorithm can be applied to arbitrary network structures and it is
computationally tractable for both small and large markets, our methodology is applicable to
wide range of empirical settings (e.g. labor markets, online buyer-seller platforms, automobile
markets, etc.).
We adapt the buyer-seller model to the labor market to explore questions about wage
dispersion and growth. In this case, workers are sellers and firms are buyers of their services.
To study wage growth, we extend our model to accommodate multiple periods. In this
extension each period has three stages. In the first stage, J new firms enter the market
4
and links are formed with the employed and unemployed workers. The parameters J and the
number of links determine the market tightness (ratio of J to unemployed workers, denoted by
θ) and the level of frictions. Firms that are employing a worker from a previous period do not
receive any new links but retain the link to their employee. In the second stage, firm-worker
matches are formed given the new network as in the basic buyer-seller model. Applying the
buyer-seller model implies that workers accept the vacancy that pays the highest wage and
hence, do not consider other aspects of the match, such as future wage growth (see section
4.4 for more on this point). Finally, at the end of the period, some matches are randomly
destroyed. The firms that are unmatched at the end of a period (either because the match
was destroyed or they could not form a match in the first place) exit the market. We interpret
“firms” as time sensitive vacancies so that, if by the end of a period, a vacancy is not filled,
it disappears from the job market. When the next period starts, J new firms enter.
The rest of the paper is organized as follows. In the next Section we describe the related
literature and highlight how our paper contributes to the current body of work. In Section 3,
we present two motivating examples. In Section 4 we describe the model. Section 5 presents
the deferred acceptance algorithm. In Section 6 we describe the results of the simulation.
Finally, in Section 7, we discuss how our results can be interpreted in the context of eBay
auctions and labor markets, and how they can be used to develop a framework for robust
econometric analysis. All proofs are in the appendix.
2 Contributions and Related Literature
Our paper contributes to the literatures on network theory, search and matching, and com-
peting auctions. Most matching market models do not consider frictions nor dynamics in
their analyses. While search models feature both dynamics and frictions, they severely re-
strict the competition between workers and firms. Competitive auction models are either
frictionless or restrict the competition between auctions.
In network theory, the closest related papers to ours are Kranton and Minehart (2001),
Corominas-Bosch (2004), Manea (2011), Gautier and Holzner (2013), Elliott (2014), and
Elliott (2015).7 These papers analyze static models where multiple buyers negotiate with
multiple sellers.8 Kranton and Minehart (2001) study efficiency when using a public ascending
auction to clear the market. They do this in two steps: first, for each exogenously given
7There is a vast literature in economics and sociology that studies information transmission in socialnetworks (for example, see Myers and Shultz 1951; Rees 1966; Montgomery 1991; Calvó-Armengol andJackson 2004, 2007; and the references there). Pairwise stability has been used to study network formation(e.g. Jackson and Wolinsky 1996). We do not study information transmission nor network formation. Incontrast, we use pairwise stability as the matchmaking criterion in a network with frictions. We characterizepairwise stable allocations and the set of prices that sustain them in networked markets. We then use thischaracterization to study price dispersion in these networks. See Jackson (2008) for a detailed review of theliterature on social and economic networks.
8Fainmesser (2012) analyzes repeated games in networks. Galeotti (2010) investigates the role of commu-nication on consumer’s search and firms’ pricing behavior in networked markets.
5
network, they study conditions under which the equilibrium outcome is efficient; second,
they study endogenous network formation and prove that the endogenously formed networks
satisfy their conditions for efficiency. Corominas-Bosch (2004) examines the equilibrium
payoffs of an alternating offers game to answer two questions: what is the set of networks that
supports a specific allocation (similar to what we call the “Walrasian outcome”), and what are
the networks that only support this allocation. Manea (2011) investigates bilateral bargaining
in networks. Gautier and Holzner (2013) study efficient allocations in arbitrary bipartite
graphs by studying the set of maximal matchings. Elliott (2014) studies the efficiency of the
entry decisions of firms and workers in a labor market model.9 Elliott (2015) extends the
Kranton and Mineheart model to consider different levels of bargaining power, different cost
shares, negotiated investments, and ex-ante heterogeneous gains from trade.
While the use of networks to model frictions in an economy is common, both the question
we seek to answer, and the way we answer it, is novel. The literature typically proposes a
concrete game to be played within these networks and focuses under which conditions (if
any) the equilibria of these games are Pareto efficient. For example, Kranton and Minehart
(2001) and Corominas-Bosch (2004) assume buyers and sellers play in simultaneous auctions
restricted by the network (a buyer can bid on a seller’s auction only if the two are linked) and
study whether the outcome of these auctions is Pareto efficient. However, to an econome-
trician, the concrete mechanism through which the goods are allocated is rarely observable.
Hence, it is useful to impose a minimal set of restrictions on this allocation mechanism. Re-
maining agnostic with respect to the details of the game allows the researcher to weaken the
behavioral assumptions that would be specific to the game otherwise imposed; for example,
whether buyers are submitting simultaneous (or sequential) bids to sellers, whether sellers
are proposing simultaneous (or sequential) prices to the buyers, whether they are alternating
bids and price propositions, etc. The weakest criterion for matchmaking that is consistent
with Pareto efficiency and agnostic regarding the game details is pairwise stability. To the
best of our knowledge, for arbitrary network structures the literature lacks a characterization
of pairwise stable matchings and their supporting prices. In this paper we characterize the
set of prices that sustains each pairwise stable match. Moreover, our focus is not on efficiency
(already embedded in the pairwise stability criterion) but, rather, on the role that network
effects play in determining the prices that support stable matchings.
9 Elliott (2014) shows that, for his particular labor market game, there exists a perfect Bayes Nash equilib-rium (PBNE) where the payoffs in a sparse network are the same as the payoffs in the perfectly competitiveoutcome. In contrast, our results are game-free: the only constraint we impose on allocations is that theybe pairwise stable. So we remain agnostic about the mechanism (game) that generates such allocations. Byimplication, we are also free of equilibrium assumptions (Nash Equilibrium, PBNE, rationalizability, etc.).The results are complementary since our model does not nest his and vice-versa. Workers in his model choosewhich firms to apply to, while workers in our model receive links exogenously. Our assumption is similarto the assumption made in the random search literature, where workers and firms meet randomly; Elliott’sassumption is similar to the assumption made in the directed search literature, where workers choose whichfirms to meet (see Rogerson, Shimer, and Wright 2005 for more on modeling decisions in search models).We have workers and firms meet exogenously since our goal is not to develop a theory of labor marketparticipation, but a parsimonious model of wage dispersion.
6
Our paper contributes to the literature on the matching role of markets (e.g. Gale and
Shapley 1962; Shapley and Scarf 1974; Crawford and Knower 1981; Kelso and Crawford
1982; Ausubel and Milgrom 2002; and Hatfield and Kojima 2010).10 We follow the matching
literature by developing a deferred acceptance algorithm that picks specific stable matchings.
The algorithm has two stages. The first stage outputs an allocation and is motivated by
the wage adjusting process in Crawford and Knower (1981) and Kelso and Crawford (1982).
This allocation has the property that there exist prices for which it is pairwise stable. The
second stage outputs two prices: the pointwise minimum price at which the stage 1 allocation
is stable, and the pointwise maximum price at which the stage 1 allocation is stable.11
There is an extensive literature in industrial organization that uses models of search to
rationalize price dispersion observed in real world markets.12 Some of these models include
fixed sample and sequential search (e.g. Stigler 1961; Rothschild 1973; Reinganum 1979;
MacMinn 1980; Burdett and Judd 1983; Carlson and McAfee 1983; Stahl 1989; Janssen
and Moraga-González 2004; Janssen, Moraga-Gonzalez, and Wildenbeest 2005; Lester 2011)
and clearinghouse models (e.g. Salop and Stiglitz 1977; Rosenthal 1980; Varian 1980; Baye
and Morgan 2001; Baye, Morgan, and Scholten 2004). These models typically do not use
networks to analyze price dispersion. In contrast, it is critical to our analysis to incorporate
network effects.
We also contribute to search models with on-the-job search. In wage posting and com-
petitive search models (e.g. Burdett and Mortensen 1998 and Moen 1997), there is ex-ante
competition among firms for workers. Firms must post and commit to wages before coming
into contact with workers. Models with Bertrand competition between firms (e.g. Postel-
Vinay and Robin 2002) have ex-post competition between firms by allowing a firm to make
counter-offers to their employees when they come into contact with a rival firm. In contrast to
this literature, we study prices and allocations in a network where we allow firms to negotiate
ex-post with more than one worker at a time.
Our model also relates to the literature on directed search with multiple applications (e.g.
Kranton and Minehart 2001; Albrecht, Gautier, and Vroman 2006; Kircher 2009; Galenianos
and Kircher 2009; Walthoff 2012; Albrecht, Gautier, and Vroman 2014). In these models,
firms also post wages and workers can apply to more than one vacancy. Since firms post
wages, workers are not able to negotiate wages ex-post between different firms. A central
question analyzed in this literature is whether the level of entry is efficient. Although an
important question, the efficiency of the entry decision is not the focus in our paper. In
Albrecht, Gautier, and Vroman (2014), workers post selling mechanisms and firms choose
one worker to interact with. They allow for wage competition between firms but not between
workers. In contrast to this literature, we allow firms to interact with multiple workers.
10Roth (2008) discusses recent progress in the study of deferred acceptance algorithms. See Roth andSotomayor (1990) for a comprehensive survey of the two-sided matching literature.
11See section 5 for details.12See Baye, Morgan, and Scholten (2006) for a detailed survey.
7
We also contribute to the competitive auctions literature where buyers are typically al-
lowed to bid in only one auction (e.g. Wolinsky 1988; McAfee 1993; Peters and Severinov
1997, 2006; Julien, Kennes, and King 2000). By allowing buyers to be linked to many sellers,
our model generates competition among sellers typically absent in most competitive auctions
models.
There is also a growing literature that uses networks to study trading in financial settings
such as over-the-counter (OTC) markets (e.g. Gofman 2011; Malamud and Rostek 2013;
Babus and Kondor 2013; and Alvarez and Barlevy 2014). They use concrete games to
investigate OTC markets where dealers trade with other dealers. In contrast, we study
markets where the set of sellers and buyers belong to two disjoint sets: sellers can only trade
with buyers while buyers can only trade with sellers (i.e. bipartite networks as defined in
Section 4).
In the computer science literature, Kakade, Kearns, and Ortiz (2004) study trade using
an Arrow-Debreu economy (without firms) where consumers trade goods with other con-
sumers. Kakade, Kearns, Ortiz, Pemantle, and Suri (2004) use a concrete game to study the
interaction between the statistical structure of the underlying network and the variation in
prices at equilibrium.13
3 Two Motivating Examples
The following two examples illustrate indirect competition and the usefulness of our theoreti-
cal tool for characterizing pairwise stable matches in networks. In both examples, we assume
that sellers are selling identical goods.
We use Example 1 to get intuition about indirect competition and network effects. In-
direct competition is a feature of the structure of the network, whereby buyers that are not
connected to the same seller have to compete with each other. Likewise, indirect competi-
tion between sellers occurs when sellers that are not connected to the same buyer have to
compete. Network effects are defined as the effect of indirect competition on prices. The
simplest example that includes these features involves two sellers and three buyers.
13Kakade, Kearns, Ortiz, Pemantle, and Suri (2004) analyze networked markets where the numbers ofbuyers and sellers are equal. They show that, for their particular game, there is no equilibrium price dispersionwhen the following conditions hold: (1) the number of buyers and sellers go to infinity, (2) the links are formeduniformly at random, and (3) the probability of forming a link is high enough. In their model there is limitingprice dispersion (as the number of buyers and sellers go to infinity) when the network is formed via preferentialattachment. In contrast, the only constraint we impose on allocations is that they be pairwise stable. So ourresults are game-free as emphasized in footnote 9. In addition, in our simulations we study price dispersionin bipartite networks varying arbitrarily the number of buyers, the number of sellers, and the number of linksper seller or buyer.
8
Example 1. Assume that buyers A, B and C are ordered in their valuations
(νA > νB > νC > 0) and sellers 1 and 2 have the same valuation (normalized to
0). Consider the following network:
1
2
A
B
C
Buyers
Sellers
The buyer-preferred stable match is for buyer A to pay νC to seller 1 and buyer B
to pay νC to seller 2. Buyer B cannot pay less than νC because buyer C will poach
seller 2. Likewise, buyer A cannot pay less than νC because buyer B will poach
seller 1. In this example, buyer A is indirectly competing with buyer C. Network
effects force buyer A to pay νC even though buyer C is not linked to seller 1. If
buyer C dropped out of the market, then both buyer A and buyer B paying zero
is the buyer-preferred pairwise stable match.
Example 2 demonstrates how we use abstractions to highlight the importance of indirect
competition and characterize pairwise stable matches. An abstraction in fully connected
networks is a decomposition of a network into fully connected subnetworks that satisfy the
following properties: (1) each node in the abstraction is a subnetwork of the original network,
(2) each link in the original network is either a link within a subnetwork in the abstraction or
a link that connects two distinct nodes in the abstraction. This construction uses that fully
connected subnetworks are competitive markets with a unique price. The following example
demonstrates one possible abstraction of a network.
9
Example 2. Consider a market with three sellers and five buyers with the follow-
ing network and an associated abstraction in fully connected networks. Assume
that the five buyers are ordered in their valuations (νA > νB > νC > νD > νE > b)
and the three sellers have the same valuation (i.e. b(1) = b(2) = b(3) = b).
An Abstraction in
Network Fully Connected Subnetworks
1
2
3
A
B
C
D
E
BuyersSellers
3
2
1
B
E
A
D
C
G
G0
G00
Even though many prices sustain it, there is a unique pairwise stable match:
Buyer B buys from seller 3, buyer A buys from seller 2, and buyer C buys from
seller 1.
Abstractions are useful to highlight how indirect competition affects price formation.14 We
14 One way to construct an abstraction is to follow three steps: (1) Form a subnetwork around each stablematch, (2) add the unmatched buyers (sellers) to one of the subnetworks that contains a seller (buyer) towhich they are linked, and (3) form a directed link between subnetworks if there is a buyer in one subnetworkthat is connected to a seller in another subnetwork. The direction of the link will point from the subnetworkthat contains the buyer to the subnetwork that contains the seller. Although there is not a unique assignmentin step 2, any assignment will characterize the same set of pairwise stable matches and their supporting prices.
Step 1 Step 2
3
2
1
B
A
C
G
G0
G00
3
2
1
B
E
A
D
C
G
G0
G00
10
proceed with five observations: (1) Consider subnetwork G’ as an independent subnetwork.
In this case, a stable match will be seller 2 selling to buyer A at a price that is greater than
buyer D’s willingness to pay. (2) Since buyer D is linked to seller 3, buyer D’s willingness
to pay to seller 2 is less than or equal to the price that prevails in subnetwork G. (3) Thus,
A is matched with seller 2 at a price greater than or equal to the price in subnetwork G,
even though A is not directly linked to seller 3. (4) We call this an abstraction, because
the identity of the buyer in subnetwork G’ linked to the seller in subnetwork G is irrelevant.
(5) Finally, the direction of the link between subnetwork G and subnetwork G’ describes the
relationship of the prices that prevails in each subnetwork, namely p(G0) < p(G).
Through the use of abstractions, our main theorem helps us understand pairwise stable
matchings and their supporting prices. We use abstractions to decompose these prices into a
competitive component and an outside-option component. We can do this because abstrac-
tions allow us to ignore information about the network that is irrelevant for calculating the
prices that sustain a pairwise stable match. Using abstractions and our main theorem, we
characterize the set of prices that sustain any given stable matching. Finally, abstractions
and our main theorem allows us to prove that our algorithm picks a pairwise stable matching.
We build the theory in the next section and discuss these points in detail in subsection 4.3.
4 The Model
4.1 Buyer-Seller Model
We consider buyers and sellers that wish to match pairwise. Sellers differ in their valuation
and offer a homogeneous good. For simplicity, we assume that sellers have no idiosyncratic
preferences over the buyer they sell to. Buyers differ in their valuation and have single unit
demand. A buyer with valuation ν that buys from a seller at price p has utility ν p and 0
otherwise. The seller’s utility is the price, p, if they sell the good, and their valuation, b, if
they do not. Single unit demand implies we focus on pairwise matching.
Matching takes place in exogenous buyer-seller networks. Each buyer is linked with a
subset of sellers.That a buyer is not (necessarily) linked to all possible sellers captures search
frictions in the environment.
The formal model we use to capture these interactions is a graph-theoretic model. A
graph is a set of nodes connected by links (or edges). We say the graph is undirected if the
direction of the link does not matter. We say that the graph is bipartite if the set of nodes
can be partitioned into two sets such that no two nodes in the same set are connected to each
other. In our framework, buyers and sellers constitute a bipartite undirected graph: first,
the set of nodes is partitioned into buyers and sellers; second, a buyer is linked to a seller
if and only if that seller is linked to that buyer; and third, no buyer (respectively seller) is
connected to another buyer (respectively seller). We formalize this in the next definition.
11
Definition (graph). Given a set V of nodes and a set E V 2 of edges we say (V,E) is
a graph. Moreover,
• We say a graph (V,E) is trivial if E = ; and V is a singleton.
• We say a graph (V,E) is finite if E is finite.
• We say the graph is undirected when, for each v, v0 2 V , (v, v0) 2 E if and only if
(v0, v) 2 E.
• We say (V,E) is a bipartite graph if there exists two disjoint sets, V1, V2 V , such
that V = V1 [ V2 and (v, v0) 2 E only if v 2 Vi ) v0 2 Vj, for i 6= j. We write these
graphs explicitly as (V1, V2, E).
• We say a bipartite graph (V1, V2, E) is fully connected if for each v1 2 V1, (v1, v2) 2 E
for each v2 2 V2.
Since graphs tell us which buyers are connected to which sellers, but they do not tell us
the valuation of buyers nor the valuation of the sellers, we extend the definition of the graph
to the definition of a network. Intuitively, a network is a graph where each node is given a
numerical value. This value is interpreted as the “valuation” of the buyer or seller. We define
price functions for networks as functions that map every possible buyer-seller edge into a
real number. This real number is interpreted as the price that would prevail if the buyer
was to buy the good from the seller. The price function is individually rational if, for each
buyer-seller edge, it specifies a price that lies between the seller’s valuation and the buyer’s
valuation. For the rest of the paper, even if not explicitly mentioned, J denotes the set of
buyers, and j indexes buyers. Similarly, I denotes the set of sellers, and i indexes sellers.
Definition (networks and prices). Let (J , I, E) be an undirected bipartite graph, and
M E be any subset of edges. Let ν : J ! R , b : I ! R, and pM : M ! R be functions
such that pM((j, i)) = pM((i, j)) for each (j, i) 2 M . We say N = (J , I, E; ν, b) is a buyer-
seller network. We say the function pM is a price function . We say pM is individually
rational (IR) if for each (j, i) 2 M , pM(j, i) 2 [b(i), ν(j)].
Given a network (J , I, E; ν, b), a matching M is any subset M E such that three
properties hold: first, each buyer is matched to at most one seller (recall that buyers have
unit demand); second, each seller is matched to at most one buyer; and finally, if a seller
is matched to a buyer then the buyer is matched to the seller. That is, if (j, i), (j, i0) 2 M ,
then i = i0; if (j, i), (j0, i) 2 M , then j = j0; and (j, i) 2 M if and only if (i, j) 2 M . We
then say an edge (j, i) 2 M is a match or, alternatively, that i and j are matched. Finally,
given a matching M , we define i : J ! I [ ; as the function that maps each buyer to
the seller with whom it is matched, or to the symbol ; if the buyer is unmatched. Likewise,
j : I ! J [; is the function that maps each seller to the buyer with whom it is matched,
12
or to the symbol ; if the seller is unmatched. Finally, given a matching M and a price
function pM , the function v summarizes the virtual price each agent (buyer or seller) pays or
is getting payed, without having to explicitly distinguish if the they are matched or not. We
call these functions v the payment functions. In symbols: for each j and i,
v(j) =
(
ν(j) if i(j) = ;
pM(i(j), j) if i(j) 6= ;,
and
v(i) =
(
b(i) if j(i) = ;
pM(i, j(i)) if j(i) 6= ;.
Next, we define pairwise stability of a matching M with respect to a price function pM .
Pairwise stability means that the edges in M are priced such that individual rationality holds,
and there are no mutually benefit matches by agents that are linked but are not matched
(i.e. agents linked by an edge e 2 E \ M ). In other words, any extension of pM to all
edges cannot yield Pareto improvements over the match M executed at prices pM . Note
that pairwise stability only requires that an agent is able to observe the prices of his linked
counterparts, but not who they are linked to.
Definition (block). Let M be a matching and pM : M ! R. We say an edge (i, j) 2 E \M
blocks (M, pM) if v(i) < v(j).
Definition (pairwise stability). Given a non-trivial network (J , I, E; ν, b) and a matching
M E, we say M is pairwise stable at prices pM if the following hold:
• No blocking: no edge (i, j) 2 E \M blocks (M, pM).
• Individual rationality: pM(i, j) 2 [b(i), ν(j)] for all (i, j) 2 M .
In fully connected networks it is simple to characterize stable matchings. Indeed, if M
is stable with respect to a price function p, then all prices must be the same. To see this,
assume i is matched to j, i0 is matched to j0, and let p be any individually rational extension
of pM to E. Then, p(j, i) p(j, i0) p(j0, i0) p(j0, i) p(j, i), where all these terms
are well defined because the network is fully connected. As a corollary, all stable matchings
can be characterized by whether there are more buyers than sellers or vice versa. Intuitively,
stable matchings are those matchings which are maximal and can be sustained by individually
rational prices that price out the side of the market (sellers or buyers) that is in excess. In this
regard, the matchings and prices we obtain from pairwise stability in fully connected graphs
are those that would prevail if this was a perfectly competitive economy. We summarize this
in the following remark.
Remark 1. Let (J , I, E; ν, b) be a fully connected network, with J = #J , I = #I. Assume
that b = maxb(i) : i 2 I minν(j) : j 2 J = ν. Let M E be a matching.
13
• If I > J , M is stable if, and only if,
– All buyers are matched: For each j 2 J there is i 2 I such that (j, i) 2 M .
– Only lowest valuation sellers are matched: If i 2 I is such that #i0 : b(i) >
b(i0) J then there is no j 2 J such that (j, i) 2 M .
– Seller valuations determine matching prices: For each (j, i) 2 E, p(j, i) = p where
p 2 [maxb(i) : (9j 2 J ) such that (j, i) 2 M, minb(i) : (@j 2 J ) such that (j, i) 2
M].
• If I = J , M is stable if, and only if,
– All buyers are matched: For each j 2 J there is i 2 I such that (j, i) 2 M .
– All sellers are matched: For each i 2 I there is j 2 J such that (j, i) 2 M .
– Sellers sell at an intermediate price: For each (j, i) 2 E, p(j, i) = p where p 2
[b, ν].
• If I < J , M is stable if, and only if,
– Only highest valuation buyers are matched: For each j 2 J if j0 : ν(j0) > ν(j)
I then there is no i 2 I such that (j, i) 2 M .
– All sellers are matched: For each i 2 I there is j 2 J such that (j, i) 2 M .
– Buyer valuations determine matching prices: For each (j, i) 2 E, p(j, i) = p where
p 2 [maxν(j) : (@i 2 I) such that (j, i) 2 M,minν(j) : (9i 2 I) such that (j, i) 2
M].
4.2 A Theorem
To characterize the set of pairwise stable matchings in a network and identify the set of
prices that can sustain them, it is convenient to abstract away from certain links in the
original graph and retain only the links that are essential for characterizing these matchings.
This yields economic insight into how these prices are determined. Our next definition, the
abstraction of a graph, identifies links of any given network that are essential for determining
prices that sustain pairwise stable matchings. Slightly abusing notation (see remark 2 below)
an abstraction of a graph is a directed graph with nodes and edges as follows: each node is a
subgraph of the original graph, and each edge in the original graph is either (i) an edge within
a subgraph in the abstraction or (ii) connecting two distinct nodes in the abstraction. We
now present a formal definition for the abstraction of a graph, an extension of that definition
for networks, and an example.
Definition (graph abstraction). Let (J , I, E) be a buyer-seller graph. We say a directed
graph (G, E) is an abstraction of (J , I, E) if the following hold:
14
• Each G 2 G is a graph (JG, IG, EG) such that JG J , IG I,
EG = e : e 2 E and e 2 (JG IG) [ (IG JG),
• JG : G 2 G and IG : G 2 G are a partition of J and I respectively,
• (G,G0) 2 E if and only if there exists j 2 JG, i 2 IG0 such that (j, i) 2 E.
Moreover, we say that an abstraction is an abstraction in fully connected graphs if each G 2 G
is a fully connected graph.
Remark 2. Generally, nodes in graphs are the smallest object in the graph. As such, a
more standard definition of abstraction would let the vertices G 2 G be arbitrary objects in
an arbitrary set, and would include a bijective mapping between G and the relevant subgraphs
of (J , I, E). However, this would imply adding notation that makes the model unnecessarily
cumbersome.
Remark 3. Since one-to-one graphs and trivial graphs are fully connected, an abstraction in
fully connected graphs always exists.
Remark 4. Given a graph (J , I, E), abstractions in fully connected graphs will not neces-
sarily be unique. See example 2 above.
Definition (network abstraction). Let (J , I, E; ν, b) be a buyer-seller network. We say
(G, E; p) is an abstraction of (J , I, E; ν, b) if the following hold:
• (G, E) is an abstraction of (J , I, E),
• If G 2 G satisfies JG = j, IG = ;, then p(G) = ν(j); if G 2 G satisfies JG = ;,
IG = i, then p(G) = b(i); otherwise, p(G) 2 [minb(i) : i 2 IG,maxν(j) : j 2
JG].
Since our objective is to use abstractions to characterize the stable matchings in a graph,
our next definition specifies when a matching is stable with respect to an abstraction of that
graph. Intuitively, a matching M E is stable with respect to an abstraction when three
conditions hold. First, the abstraction does not break M : for each buyer-seller match in M ,
that pair belongs to the same subgraph in the abstraction. Second, recall that the only price
functions that can sustain a stable matching in a fully connected network are those where
all edges are priced equally. Therefore, each subgraph G in the abstraction is assigned a
number, p(G), that plays the role of this uniform price. The third condition is that buyers
are sorted into the fully connected subnetwork whose price is lower than the price of any
other subnetwork they have access to.
Definition (stability abstraction). Let (J , I, E; ν, b) be a buyer-seller network and (G, E; p)
be an abstraction of it in fully connected graphs. We say that M is stable with respect to the
abstraction if three conditions hold:
15
• G does not break M : for each e 2 M , e 2 EG for some G 2 G.
• Prices p(·) induce pairwise stability:
– For each non-trivial G 2 G, M restricted to G is stable at prices p(j, i) = p(G)
for all (j, i) 2 M \ EG,
– If G = (i, ;, ;) for some j, then p(G) = b(i),
– If G = (;, j, ;) for some j, then p(G) = ν(j).
• Cheapest sorting: if (G,G0) 2 E then p(G) p(G0).
Remark 5. The cheapest sorting condition is defined to be consistent with the construction
of the edges in E. Indeed, we define E so that (G,G0) 2 E whenever a buyer in G is linked
to a seller in G0. Since buyers search for the cheapest price, then we define cheapest sorting
as p(G) p(G0). If we took the opposite convention for the edges in E, that (G,G0) 2 E
whenever a seller in G is linked to a buyer in G0, since sellers search for the highest price,
then cheapest sorting would be defined as p(G) p(G0).
With these definitions we can state our main theorem.
Theorem 1. Let N = (J , I, E; ν, b) be a network. Let M be a matching. Then the following
are equivalent:
1. There exists pM such that M is stable with respect to pM
2. There exists an abstraction in fully connected graphs A = (G, E; p) such that M is
stable with respect to A.
The proof of the theorem is in section A in the appendix.
4.3 Theorem Application
In this subsection we illustrate four ways in which abstractions and Theorem 1 are useful to
understand pairwise stable matchings and their supporting prices. For this we use Example 2.
First, theorem 1 is useful to decompose the prices that sustain pairwise stable matchings
into a competitive component and an outside option component. We proceed in four steps.
First, consider nodes G’ and G as independent graphs and note that they are fully connected
graphs. Second, use Remark 1 to conclude three things: that only buyers B and A should
match to sellers, that buyer B should pay at least ν(E), and that buyer A should pay at least
ν(D). Steps one and two imply that we would observe these matches and supporting prices
if these were two independent, perfectly competitive economies. The later two constraints
are the competitive component of prices that sustain pairwise stable matchings. Third, note
that buyer D is linked to seller 3, as indicated by the edge (G0, G) in the abstraction. Thus,
any price that sustains a pairwise stable matching must also satisfy that buyer A pays seller
16
2 no more than what B pays seller 3. This is reflected by the cheapest sorting condition, and
is what we call the the outside option component. Steps one through three imply that all
prices that support a pairwise matching that is stable with respect to an abstraction will have
these competitive and outside option components. Fourth, Theorem 1 implies that all such
prices make this matching stable with respect to the original network. Therefore, all prices
that support pairwise stable matchings (in the original network) also have these competitive
and outside option components.
A corollary of the above decomposition is that, to calculate the prices that sustain a
particular stable matching, not all edges are relevant. Indeed, consider a modified graph
where edge (D, 3) is replaced by edge (A, 3). The original matching is still stable in the
abstraction at the original prices. Thus, this matching is also stable in the modified graph.
The relevant aspect of these graphs that sustains the proposed matching at the proposed
prices is that at least one buyer in A,D is connected to seller 3, but the exact identity
of the buyer is irrelevant. Conversely, given an abstraction and a pairwise stable matching
in that abstraction, any graph obtained by drawing edges in a manner consistent with the
abstraction will support the given matching.
Second, Theorem 1 is also useful to pin down the set of all prices that can sustain any given
stable matching. In Example 2, the unique stable matching is M = (C, 1), (A, 2), (B, 3),
but there are many prices that can sustain M . To calculate the full set of prices that sustain
M we consider the abstraction in fully connected graphs shown in Example 2: in G, the unique
stable match is (B, 3) at a price p(B, 3) 2 [ν(E), ν(B)]; in G0 the unique stable match is
(A, 2) at a price p(A, 2) 2 [ν(D), ν(A)]; and in G00 the unique stable match is (C, 1) at
a price p(C, 1) 2 [b, ν(D)]. However, since G00 is connected to G0 and G0 is connected to
G, we must also have p(C, 1) p(A, 2) p(B, 3). Therefore, M can only be sustained at
prices that satisfy p(C, 1) 2 [b, ν(c)], p(A, 2) 2 [maxp(C, 1), ν(D),minp(B, 3), ν(A)] =
[maxp(C, 1), ν(D), p(B, 3)], p(B, 3) 2 [maxν(E), p(A, 2), ν(B)] = [p(A, 2), ν(B)].
Third Theorem 1 is also useful to prove that the algorithm (see Section 5) finds: (1)
pairwise stable matchings in any given network and (2) the upper and lower bounds of the
set of prices that sustain those matchings. In Section 5 we present a description of the
algorithm and its properties. In Appendix B we present the formal algorithm and formal
proofs.
Finally, Theorem 1 allows us to rationalize pairwise stable matchings as the result of
equilibrium bidding strategies in simultaneous auction games. In the example above, we
can interpret each node in the abstraction as the following second-price auction. Each seller
holds an auction. Buyers can only bid in the sellers’ auction that belongs to same node
in the abstraction. In the example, buyer A only bids in seller 2’s auction, buyer B only
bids in seller 3’s auction, buyer C only bids in seller 1’s auction, and so on. Finally, we
assign fictitious values to the buyer with the second highest valuation in each node of the
abstraction. This fictitious value is determined by the constraints imposed by the edges in
17
the abstraction. In the example, C has fictitious value ν(C) = p(C, 1); D has fictitious
value ν(D) = p(D, 2) 2 [maxp(C, 1), ν(D), p]; and E has fictitious value ν(E) = p(B, 3) 2
[p(A, 2), ν(B)]. Such triplets (p(C, 1), p(A, 2), p(B, 3)) make the matching M stable with
respect to the abstraction. Alternatively, instead of assigning fictitious values to the buyers
with the second highest valuation, we can assign reservation prices to the sellers. Again, these
reservation prices are determined by the constraints imposed by the edges in the abstraction.
The outcome of the independent second-price auctions with fictitious valuations for the buyers
is observationally equivalent to the independent second-price auctions with reservation prices
for the sellers. Moreover, the outcome of either of these independent second-price auctions is
indistinguishable from the outcome of any other mechanism that generates the same matching
at the supporting prices.
4.4 Labor Market Model with On-the-job Search
In this section we adapt the buyer-seller model to the labor market, where workers are sellers
and firms are buyers. We assume workers do on-the-job search and that firms have single
unit demand, so a firm is equivalent to a vacancy.
At the beginning of period 1, a finite bipartite graph is randomly drawn between workers
and firms. We use (I1,J1, E1) to denote the time 1 graph, I and J to denote the respective
number of workers and firms, and we define market tightness as the ratio of J to unemployed
workers, denoted as θ1. Conditional on the graph we assign productivities to firms, denoted
with ν(·), drawn i.i.d. from a continuous distribution F with support in the interval [ν, ν].
To keep the model simple we assume that all workers have the same reservation wage. Using
the notation of the buyer-seller model, b(i) = b for each i 2 I1.
After the period 1 graph has been realized, we pick a specific pairwise stable matching
M01 at wages wM0
1using the algorithm that we describe in Section 5. Matched firms receive a
period utility of ν(j)wM0
1(j, i(j)), matched workers receive a period utility of wM0
1(j(i), i),
unmatched workers receive their reservation wage b, and unmatched firms receive utility 0
and leave. After these utilities are realized, there is an exogenous job destruction shock. This
means that each link m 2 M01 is dissolved with probability δ > 0. Firms whose links are
dissolved become unmatched and leave. We denote with M1 the period 1 matching after the
exogenous job destruction.
Now consider the beginning of period t 2. Given the matching Mt1 of period t 1,
we add J new firms and no new workers (i.e. It = It1). We draw the productivities for
these new firms from the same distribution F . We randomly draw links between workers
and firms that satisfy the following three conditions. First, positive probability is assigned
only to graphs with vertices in It [ Jt, where these denote the set of period t workers and
firms, respectively. Second, matches from period t 1 are not dissolved (that is, Mt1 Et,
where Et is the set of period t edges). Finally, matched firms receive no new applications, but
matched workers may apply to new firms because they can do on-the-job search (that is, if
18
(j, i) 2 Mt1 then i0 : (j, i0) 2 Et = i, but no constraints are placed on j0 : (j0, i) 2 Et).
We denote the corresponding graph with (It,Jt, Et). The reservation wage of workers who
were not matched in t1 is b; the reservation wage for workers who were matched in t1 is the
worker’s wage (wMt1). As before, the period utility for matched firms are their profits, the
period utility for matched workers are their wages, the period utility of unmatched workers
are their reservation wages, and the period utility of unmatched firms are 0 and these leave.
Finally, period t utilities are discounted at a rate βt, with β 2 (0, 1).
We are applying the buyer-seller model within each period, so pairwise stable matchings
are independently formed period by period. Determining the matches in this way implies
that workers accept the vacancy that pays the highest wage and hence, do not consider other
aspects of the match, such as future wage growth. From the firm’s perspective this is without
loss of generality: if they are unmatched at the end of a period they leave, so their static
and dynamic problems coincide. From the workers perspective, however, there is a loss of
generality. To see this consider worker i that is matched to firm j at wage w at the end of
period t, and assume that in period t + 1 firm j0 will only be linked with i. To rule out the
trivial case, assume that ν(j0) > w. Then worker i will expect a wage in period t + 1 that
belongs to [minν(j), ν(j0),maxν(j), ν(j0)]. Hence, worker i is willing to work for a more
productive firm in period t even if that firm offers slightly lower salaries than the competitors.
To the best of our knowledge, there is no standard solution concept for dynamic matching
markets when matching opportunities arrive over time. For a more thorough discussion of
the complications that arise in dynamic matching models see, for example, Doval (2014). For
this reason, relaxing this assumption is left for future work.
5 A Deferred Acceptance Algorithm
We now present our deferred acceptance algorithm. We describe the algorithm as a first-price
auction to give intuition of how the algorithm works. A formal description of the algorithm
can be found in Section B in the appendix. We denote the agents on the side of the market
that are holding the “auctions” as sellers and the agents on the other side that are “bidding”
as bidders. Recall that we are approaching this problem from the matching perspective, so
we are not making any statement about the actual economic mechanisms or incentives of the
agents that determine prices and matches. Bidders bid in increments of ∆
2. The value of ∆
is set so that the productivity of firms lie in a ∆ grid. Formally, for all j, ν(j) = b + kj∆
for some integer kj that is randomly drawn at the start of the algorithm. We describe the
algorithm for the case where the sellers hold the auctions. When buyers hold the auctions,
the bidding starts at their valuation and prices decrease.
The algorithm has two stages. The first stage outputs an allocation and is motivated by
the wage adjusting process in Crawford and Knower (1981) and Kelso and Crawford (1982).
(See Section B in the appendix for a detailed comparison about the first stage of our algorithm
19
and the algorithms in Crawford and Knower and Kelso and Crawford.) This allocation has
the property that there exist prices for which it is pairwise stable. The second stage outputs
two prices: the pointwise minimum price at which the stage 1 allocation is stable, and the
pointwise maximum price at which the stage 1 allocation is stable.
Stage 1: The Matching Determination Program
The algorithm starts in round t = 1 when none of the sellers has received any bid. All
bidders are placed into a queue and arrive sequentially. The entering order of the bidders is
determined randomly. The standing bid of a seller is the last bid accepted by the seller or b
if the seller has not received any bids. The winning bidder is the bidder who placed the last
standing bid.
This is round t of the matching determination program.
1. Take the first bidder in the queue (for concreteness, call it bidder j). Bidder j selects
the seller with the lowest standing bid among the linked sellers. If there is more than
one such seller, the bidder selects one of these sellers at random. Call it seller i. If
the lowest standing bid is greater than ν(j) ∆
2, bidder j does nothing and leaves the
queue. Otherwise, bidder j bids the standing bid of seller i plus ∆
2.
2. If bidder j makes a bid, seller i accepts the bid from bidder j. The new standing bid
of seller i is now the previous standing bid plus ∆
2. Bidder j leaves the queue. If there
was a bidder j0 who was the winning bidder (before bidder j bid), bidder j0 is placed
at the end of the queue.
3. The algorithm continues from step 1 with the next bidder in the queue. The algorithm
stops when there are no bidders left in the queue. In this case, each seller is matched
to the winning bidder.
We now present the second stage, the price determination program. The key insight of
this stage is that, if a seller i is matched to a buyer j, and is also linked to an unmatched
buyer j0, then the price j pays i must price j0 out of the market. That is, pM(i, j) ν(j).
Moreover, if seller i is matched to buyer j, and seller i is also linked to a buyer j0 who is also
matched (say, to a seller i0) then i must be getting payed at least what i0 is getting payed.
Otherwise, j0 would like to block with i.
Stage 2: The Price Determination Program (I)
The program starts in round t = 1 with M E produced from stage 1 as its input.
1. Set the “price” of all unmatched sellers to b.
20
2. For matched sellers, set the price of seller i for buyer j to the maximum ν(j0) amongst
all j0 that are linked to i but are not matched.
3. We call these prices (ρ1i )i2I .
This is round t > 1 of the price determination algorithm. We take (ρt1i )i2I as inputs for
this round.
1. Set the “price” of all unmatched sellers in round t to b.
2. For matched sellers, set the price of each seller i for buyer j to the maximum price in
round t 1 of the matched buyers that are linked to i. That is, amongst all matched
j0 that are linked to i, set ρti to the maximum ρt1i(j0). Note that one such j0 is j itself,
so these prices form a non-decreasing sequence.
3. If ρti = ρt1i for all i, stop the algorithm and output these prices. Otherwise, start step
t+ 1.
As formally stated in Proposition 1, the Price Determination Program (I) captures the
pointwise minimum price function at which M is stable. A modified version of this program,
which we call Price Determination Program (II), generates the pointwise maximum price
function at which M is stable. Rather than starting with ρ1 at a low value, with successive
iterations iterations rising it, the modified program starts with ρ1 at high values and successive
iterations lower it. Section B contains the formal algorithm, including both versions of the
Price Determination Program.
Proposition 1. The deferred acceptance algorithm has the following properties:
1. It stops after a finite number of rounds.
2. It outputs a pairwise stable allocation.
3. Price Determination program (I) outputs the pointwise minimum price function at
which M is stable.
4. Price Determination program (II) outputs the pointwise maximum price function at
which M is stable.
The proof of Proposition 1 is in section B in the appendix.
6 Results
In the next two subsections we document the results from simulations of the buyer-seller and
the labor-market models. We use the results from the simulation to obtain predictions about
the population distribution of prices and the matching process.
21
6.1 Buyer-Seller Model
6.1.1 Simulation
We now describe the simulation of the buyer-seller model.
There are three parameters in the buyer-seller model: the number of buyers (J), the
number of sellers (I), and the expected number of links per buyer (ELB). Every seller begins
with one unit of a good (so the number of goods is I). The market tightness, θ, is the ratio
of the number of buyers to the number of sellers, θ = JI. The market tightness is exogenous.
We start the baseline simulation with I = 10, 000 identical sellers and J = 10, 000 θ
heterogeneous buyers.15 We consider markets with J 2 [1000, 50000], so θ 2 [0.1, 5]. We
also consider markets with ELB 2 [1, 10].16 The higher the ELB, the lower the search
frictions in the market. The product of the number of buyers and the ELB determines the
number of active links in the market. The total number of possible links in the market is
J I. The proportion of active links relative to the total number of possible links in a
network is a measure of the sparsity of the network. Given the parameters J , I, and ELB, a
network is formed by randomly drawing buyers and sellers to form links. Once the network
is constructed, we apply the algorithm from Section 5 to the network. The “bids” in the first
stage of the algorithm take place on a grid of possible prices with 2J grid points.
Buyers’ valuation is normalized to range between 0 and 100 which bounds the minimum
and maximum prices between those values. One can interpret the reported prices as if
the buyers’ valuations are drawn from a uniform distribution, whose support is normalized
between 0 and 100. Alternatively, one can interpret the prices or valuations as percentiles of
any cumulative distribution function of the buyers’ valuations. The second interpretation is
possible because the allocations and prices only depend on order statistics and not the actual
valuations.
We compare the price distributions to the Walrasian outcome, when each buyer is linked
to every seller. The Walrasian outcome price, pwalras, is given by:
pwalras =
8
<
:
0 if θ 1
(1 1θ) 100 if θ > 1.
Recall that the Walrasian outcome has a unique price (see Remark 1). When θ 1, there
are more sellers than buyers and so there is always a seller who is indifferent between selling
the good at 0 or not selling it at all. In other words, the reservation price of the marginal
seller is zero, which is what determines the market price. When θ > 1, there are more buyers
than sellers. Only 1θ
of the buyers will buy the good. Hence the valuation of the marginal
buyer will be (1 1θ) 100. This buyer will be indifferent between paying (1 1
θ) 100 and
leaving the market, and so the market price will be (1 1θ) 100.
15The results do not change substantially using 1,000 or 100,000 sellers. Results are available upon request.16We obtain similar results by varying expected links per seller (ELS) in the simulations.
22
6.1.2 Results
Distribution of Prices. Figure 1 displays the distribution of prices for the buyer-preferred
match by market tightness (horizontal axis in each panel) and ELB (different panels). Each
vertical box corresponds to a simulated market characterized by those parameters. Each
panel shows the population distribution of prices for different levels of search frictions in
different markets. The top-left panel shows the price distribution for high frictions, where
ELB equals 1. The top-right and bottom panels show what happens in markets with lower
frictions (when ELB equals 2, 3, and 5, respectively). At low levels of θ there are many sellers
for each buyer. So low numbers for θ indicate “loose” seller markets where sellers are at a
disadvantage. In addition, each panel displays the Walrasian outcome.
For market tightness less than one, the market looks like a monopsony and nearly all sellers
are paid their valuation (recall that, for simplicity, all sellers are identical, so we normalized
their valuation to zero). This is because it is unlikely for a seller to receive multiple links.
Even if a seller receives two links, it is likely that at least one of the buyers has an outside
option of zero. This happens if the buyer is also linked with another seller who has no other
links.
On the other hand, as market tightness is increased the market becomes more competitive
between buyers and more favorable for sellers. The median price increases as does price
dispersion. There are now many buyers linked to each seller and the buyers have worse
outside options. Even if a buyer is linked to a second seller, it is likely that the second seller
is linked to many other buyers. In markets with lower frictions, competition between buyers
increases, thus increasing prices until they reach the Walrasian outcome.
Figure 2 shows that similar results to the ones in Figure 1 are obtained using the seller-
preferred match. Figure 2 displays, for each market tightness, the distribution of prices using
both the seller- and the buyer-preferred match. (For the buyer-preferred match, each vertical
box in Figure 2 is identical to the corresponding vertical box in Figure 1.) When ELB equals
5, the 95th and 5th price percentiles coincide with the Walrasian outcome for both the buyer-
and the seller- preferred match. The prices in the buyer-preferred matching represents the
lower bound of the set of prices that support each match. Likewise, the prices in the seller-
preferred match represents the upper-bound of the set of prices that support each match.
Since both the seller-preferred and buyer-preferred price distributions mimic the Walrasian
outcome when ELB=5, it must be true that the price distribution in any allocation that
supports a pairwise stable match must also mimic the Walrasian outcome.
Price Dispersion and the Walrasian Outcome. Price dispersion decreases when search
frictions decrease. There are many buyers linked to each seller, but there are also many
sellers linked to each buyer, improving the outside options of both parties. These improved
outside options reduces price dispersion (i.e. the likelihood that a seller has to take a low
price is low, but at the same time the probability that a buyer has to pay a high price
23
is also low). Figure 3 shows the evolution of the price distribution for the buyer-preferred
match. In the top panel, the figure displays the difference between the 95th and the 5th
price percentiles. In the bottom panel, the figure displays the difference between the 99.5th
and 0.5th price percentiles. All sellers are paid the same price at the Walrasian outcome, so
both differences equal zero at the Walrasian outcome. We are interested in answering the
following two questions: How sparse can the network be while 90% and 99% of sellers are
paid the same price? While there is price dispersion when there are fewer than four ELB, the
price distribution begins to collapse for more ELB. When there are five ELB, there is nearly
no difference between the price at the 95th and 5th percentiles. Likewise, when there are
eight ELB, there is almost no difference between the price at the 99.5th and 0.5th percentile.
In other words, at least 90% or 99% of the sellers are paid the same price when the number
of active links relative to the total number of links is only 5/10,000 or 8/10,000, respectively.
The price distribution in the model collapses with less than 0.1% of the possible links in the
network.
The Effect of Frictions on Mean Prices. Figure 4 displays the evolution of mean
prices (where the expectation is taken relative to the population distribution of prices) over
ELB for different market tightness. Mean prices represent the buyers’ and sellers’ ex ante
expected prices before the network is drawn. The figure shows how mean prices vary with
ELB (i.e. frictions) in a given market (holding fixed the market tightness), so mean prices are
normalized by the mean price when ELB equals 1. Increasing ELB may increase or decrease
mean prices, depending on market tightness. For example, consider markets where there are
many sellers for each buyer (θ 1, so that sellers are at a disadvantage). When ELB is
low, price dispersion is high, even when there are more sellers than buyers (top-left panel in
Figure 1), resulting in relatively high mean prices. As ELB increases, the price distribution
collapses to the Walrasian outcome (Figure 3). The Walrasian outcome is zero when there
are more sellers than buyers. Thus, when there are more sellers than buyers (θ 1), lowering
frictions results in lower mean prices as a consequence of network effects. Intuitively, since
there are more sellers than buyers, increasing ELB improves the outside option of the buyers
who now talk to relatively more sellers, even when sellers expect to talk to more buyers.17
Distribution of Matched Buyers. Figure 5 shows the distribution of matched buyers
for different markets. In loose markets (θ < 1), the probability of finding a match does
not depend on the buyer’s valuation. Prices are low and there are many unmatched sellers,
so buyers have a roughly equal chance of finding a match. The ELB does not change the
distributions of matched buyers in loose markets.
As markets become tighter (θ > 1), competition between buyers becomes more important.
In these markets, prices are higher and some buyers are priced out of the market. Buyers
17Same results are obtained using expected links per seller (ELS) instead of ELB. When there are moresellers than buyers ( 1), increasing ELS results in lower mean prices. Results are available upon request.
24
with high valuations (e.g. above the Walrasian price) are more likely to buy goods than
buyers with low valuations. When ELB is low, buyers with high valuations may be linked
to a seller with another high-valuation buyer. Since they have few links, they are priced out
of the market. For markets with higher ELB, buyers have better outside options and the
probability that a high-valuation buyer is priced out of the market decreases. When ELB=5,
the distribution of matched buyers looks close to the Walrasian outcome, where all buyers
with ν(j) > pwalras are matched and all buyers with ν(j) < pwalras are priced out of the
market.
Welfare. Results on the welfare in the buyer-seller model are in the online appendix (see
Figure A1). We analyze welfare using the labor market model on page 28, where the results
are similar to the buyer-seller model.
6.2 Labor Market Model
6.2.1 Simulation
We adapt the buyer-seller model to the labor market to explore questions about wage dis-
persion and growth. In this case, workers are sellers and firms are buyers of their services.
Firms have single unit demand for labor and cannot dismiss their employee. We assume that
workers are homogeneous and firms are heterogeneous in their productivity (ν(j)).18 We
normalize the reservation wage of the worker to zero. If a worker and firm match at wage w,
the worker’s utility is w and the firm’s profit is ν(j) w. Wages and firm productivities are
normalized to range between 0 and 100 as in the buyer-seller model. To study wage growth,
we extend our model to accommodate multiple periods.
In the first period all workers start unemployed and the simulation is identical to the
basic buyer-seller model (see subsection 6.1). At the end of the period, some matches are
randomly destroyed at rate δ 2 (0, 1). The firms that are unmatched at the end of a period
(either because the match was destroyed or they could not form a match in the first place)
exit the market. We interpret “firms” as time sensitive vacancies. So if by the end of a period
a vacancy is not filled, it disappears from the job market.
At the beginning of the next period, some fraction of workers are employed by old firms
and the rest are unemployed. The same number of firms are created (J) and a new network
is drawn between all the workers (employed and unemployed) and the new firms. Firms from
previous periods maintain the link to their employed worker and lose all other links from
previous periods. New firms are placed into the queue and old firms start off as the highest
bidders in their employee’s “auction”. So from the standpoint of the algorithm, the standing
bid (or reservation wage) in an employed worker’s auction is the wage from the previous
18Worker heterogeneity is important for understanding wage dispersion in the data and it is straightforwardto add worker heterogeneity to our model. However, not including worker heterogeneity makes the expositionof our results more clear as the Walrasian outcome has a unique wage.
25
period. The new firms either “bid” for unemployed workers or try to poach employed workers
from old firms by bidding in the auctions of the employed workers. Firms from previous
periods can only bid in their employee’s auction. When firms arrive at the front of the
queue, they bid in their subnetworks according to the algorithm (see Section 5). The bidding
process ends when there are no more firms in the queue. Matched firms produce with the
hired workers and the workers receive their wage. At the end of the period matches receive
a job destruction shock at rate δ.
We consider markets that are in steady state. The market is in steady state when the
flows into unemployment equal the flows out of unemployment. We find the steady state by
simulating the economy for enough periods until the average unemployment remains stable.19
Then we record 400 periods.
The labor-market model has five parameters: the number of workers (I), which is constant
for all periods, the number of firms (J) that enter in each period, the expected number of
links per firm (ELF), the job destruction rate (δ), and the relative probability of receiving a
link between employed and unemployed workers (λ). The market tightness, θ, is defined as
the ratio of firms to unemployed workers, JU. In contrast to the basic buyer-seller model, θ is
now endogenous as it depends on the the number of unemployed workers (U).
Both the number of workers and the job destruction rate are fixed for all simulations.
Following Shimer (2012), the monthly employment to unemployment rate is set at 2% (prime
age men, Figure 3), which translates to a quarterly δ = 0.06. We use 5,000 workers for the
simulation of the dynamic labor market model.20 A market is a combination of J and ELF.
The relative probability of receiving a link between employed and unemployed workers
(λ) is an important determinant of the structure of the network. Most empirical studies
find different job offer rates between employed and unemployed workers.21 This is impor-
tant for understanding allocations and wage growth, since network effects are substantially
diminished when θ < 1 and λ = 1. To understand why, recall that when θ < 1, there are
more unemployed workers (U) than vacancies entering the market (J). Even when the un-
employment rate is relatively high (10%-20%), firms have a low probability of linking with an
unemployed worker when λ = 1. So even in markets that appear unfavorable for the workers
(high unemployment and low market tightness), it is difficult for the firms to link to unem-
ployed workers when λ = 1. This implies that in these markets, unemployed workers do not
have to compete with each other. If we follow the empirical literature by setting λ < 1, then
firms have a higher probability of linking with an unemployed worker and network effects
19We compute the average unemployment over 10 consecutive periods and compare it to the averageunemployment 10 periods before. We find the steady state when this difference is less than the tolerancelevel. The convergence is relatively fast. It takes between 30 to 100 periods (depending on the values of theparameters) to find the steady state. Let Ut be the unemployment in period t in a specific market and letUt =
110
Pt
j=t9 Uj be the average unemployment of the 10 periods ending in t. We find the steady state,
tSS , as follows: | UtSS10 UtSS|< , where is a tolerance level.
20The results do not change much when using 1,000 or 10,000 workers. Results are available upon request.21See Holzer (1987) and Blau and Robins (1990) for an investigation on different search intensities and
offer rates for employed and unemployed workers.
26
again become important even when market tightness is low.
To make the comparison to labor-search models, we use a model of Bertrand competition
between two firms as our benchmark. For example, Postel-Vinay and Robin (2002) (hence-
forth PVR) allows two firms to compete over the wages (à la Bertrand) of an employed worker
and is closest to our model. The PVR framework in terms of networks is as follows. Given
a set of firms, workers, and a matching technology à la PVR, construct the corresponding
network with the same set of firms and workers, where a worker is linked to a firm if, and
only if, the worker and the firm are matched by the matching technology. An important
assumption in most search models, including PVR, is that a firm can negotiate with at most
one worker and, in PVR, a worker can negotiate with at most two firms. This limits the type
of networks that are realized and drives most of the differences between search models and
our model.
We also compare our results to a perfectly competitive model (Walrasian outcome) in
steady state. Recall that a market is in steady state when the flows into unemployment
equal the flows out of unemployment. In the Walrasian outcome, all unemployed workers
find jobs when θ 1 (i.e. there are more vacancies than unemployed workers). So the
number of employed workers in steady state at the end of a period is the total number of
workers in the market adjusted by the separation rate, I (1 δ). When θ < 1, the fraction
of unemployed workers in steady state, uSS, is given by equalizing the flow of workers out of
and into unemployment:
uss + δ(1 uss)
θ = δ(1 uss),
where uss is the unemployment rate before the job separation shock.22 The left-hand side
represents the number of firms that hire workers (θ times the fraction of unemployed at the
beginning of a period). The right-hand side represents the number of workers that lose their
jobs. Thus, uSS = δ(1θ)(1δ)θ+δ
when θ < 1. Then, the number of employed workers in the
Walrasian outcome, ESS, is given by ESS = I (1 uss). Due to the multi-period aspect of
the model, the number of firms employing a worker in steady state, ESS, is greater than the
number of firms that enter in each period, ESS > J .
6.2.2 Results
Distribution of Wages. Following the empirical results in Bagger, Fontaine, Postel-Vinay,
and Robin (2014), we set λ = 0.05. This implies that an unemployed worker is twenty times
more likely to receive a link than an employed worker. Setting λ below one concentrates
the network between the new vacancies and the unemployed workers. Figure 6 displays the
22We use lowercase letters for rates (e.g. the unemployment rate is u) and uppercase letters for levels (e.g.the number of unemployed workers is U).
27
distribution of wages in different markets when λ = 0.05.23 (This figure is similar to Figure 1
but for the dynamic labor market model.) All workers now compete with unemployed workers
and the average wage is close to the Walrasian outcome when ELF equals five.
Worker Mobility and Wage Growth. To analyze worker mobility and wage growth we
keep λ = 0.05. Figure 7 shows the wage profiles of workers after an unemployment spell. In
loose markets, workers start out with low wages and wages grow very slowly. In tight markets
with low frictions, workers attain a high wage immediately out of unemployment and then
their wages grow very little over their career. This leads to the result that as you reduce
frictions, workers have lower median wage growth. Figure 8 (top panel) shows the median
wage growth after twenty periods of employment. As a benchmark, we also display the wage
growth for the Bertrand competition benchmark (see subsubsection 6.2.1). When θ < 1,
wage growth is reduced because firms have better outside options. When θ > 1, reducing
frictions causes firms to compete strongly for workers. This drives initial wages up leaving
little room for wage growth.
The lower panel of Figure 8 displays the worker mobility, which is defined as the fraction of
workers that make a job-to-job transition in a period. The intuition for why worker mobility
decreases as frictions are lowered is similar to wage growth. When θ < 1, firms are less likely
to poach workers from another firm because they have better outside options that likely
include an unemployed worker. When θ > 1, workers are less likely to move from one firm
to another because when they enter the labor market they immediately find a job that pays
a high wage. Our results indicate that empirical researchers should be cautious when using
worker mobility to make inference about the level of frictions in a market where network
effects are present.
Welfare. To analyze welfare we need a welfare criterion. In the labor market model,
the utility function of the firms that hire a worker is the production function minus the
wage, ν(j) w(j), and the utility function of the workers that are hired is the wage, w(j).
We use the utilitarian welfare criterion, Ω, that is sum of these utilities and corresponds to
Ω =PJ
j=1 ν(j), where j = 1, . . . , J index the firms that hire a worker. The unconstrained
first-best allocation is the one that maximizes the welfare in the absence of frictions. It
corresponds to the Walrasian outcome.24
Following Bagger, Fontaine, Postel-Vinay, and Robin (2014), we use the following cumu-
23Figure A2 in the online appendix displays the distribution of wages in different markets when = 1.As discussed in the previous section, = 1 is inconsistent with the empirical literature and suppressesnetwork effects. Comparing figures 6 and A2 shows the effect that including indirect competition has onwage dispersion and average wages.
24The same welfare analysis holds for the buyer-seller market using buyers, sellers, prices, and a singleperiod. We obtain similar results. See Figure A1 in the online appendix.
28
lative distribution function for the productivity of firms:
F (ν) = 1 exp
c1(ν c0)c2
,
where c0 = 5, c1 = 8, and c2 = 0.7. Thus, the utilitarian welfare in each market, Ω,
represents the aggregate production in that market.
Figure 9 displays the aggregate production (welfare) and average labor productivity for
different levels of frictions and for the Walrasian outcome (i.e. unconstrained first-best
allocation).25 The top panel in Figure 9 shows the ratio of the aggregate production in our
model relative to the aggregate production in the Walrasian outcome.26 As ELF is increased,
the aggregate production in our model is close to the aggregate production in the Walrasian
outcome. For example, when θ = 0.5 and ELF=5, the aggregate production is approximately
97.5% of the aggregate production in the Walrasian outcome. So even in sparse networks,
markets are close to the unconstrained first-best allocation.
We also investigate the allocation of workers to firms. The bottom panel of Figure 9
displays the average productivity of jobs of employed workers. We find that lowering frictions
can improve or worsen the allocation of workers to more productive firms. In loose markets
(lower θ), reducing frictions lowers average labor market productivity. The intuition is that
when frictions are high, firms are less likely to have an outside option when linked with
an employed worker. This leads to competition between firms. When frictions are lower,
firms are more likely to have another link to an unemployed worker. Hence, there is less
competition between firms. Recall that when there are no frictions and θ < 1, there is no
competition between firms since there are more unemployed workers than firms. In tight
markets (high θ), lower frictions leads to the allocation of workers to the more productive
firms as in standard search models.
7 Concluding Remarks
The defining characteristic of markets with frictions is that similar goods or services are traded
at different prices, resulting in price dispersion. In this paper we use networks to characterize
pairwise stable allocations and their supporting prices in buyer-seller markets with frictions.
The central insight of the paper is that including indirect competition causes markets with
frictions to have prices and allocations that look close to the Walrasian outcome. To study
prices in large networks, we develop a computationally tractable algorithm that finds the
upper and lower bounds of the set of prices that sustain any pairwise stable match. Network
effects reverse the relationship between the level of frictions and many economic outcomes.
25All specifications display the aggregate production in steady state at the end of the period after thedestruction shock occurred. So, ESS I (1 ) in our model, and ESS = I (1 ) in the Walrasianoutcome. The results do not change substantially if the welfare is calculated before the destruction shockoccurs.
26Figure A5 in the online appendix shows the total aggregate production.
29
We find that lowering frictions leads to: 1) lower wage growth, 2) lower worker mobility, and
3) lower expected prices in loose markets (θ < 1).
The main finding of our paper, that sparse networks look as if they were perfectly com-
petitive, might seem inconsistent with the price dispersion observed in eBay, labor markets,
automobile markets, etc. There are three dimensions in which our model provides a richer
understanding of frictions and price dispersion. Consider the case of eBay, where search
frictions arise because its search engine is quite sensitive to the buyers’ search inquiry and to
the sellers’ title for the product listing. In addition, buyers most likely do not compare all
the listings for a given product at a given time. First, one possible explanation for the price
dispersion is that search frictions are relatively high (i.e. ELB< 3 or buyers participate in
less than three auctions on average). But this is unlikely to be the whole answer. Second,
the structure of the network is an important factor in the formation of prices. Although our
networks are generated by randomly forming links, this is clearly not the case at eBay. All
buyers who make the same search inquiry receive the same list of items or products. Price
dispersion in eBay may be more about the structure of the network and less about the abso-
lute level of frictions. Third, our model makes clear predictions on the distribution of prices
for any network. Given information on the participation of buyers in auctions, the actual
network can be constructed. This can be used to decompose the sources of price dispersion
into frictions and other factors, such as unobserved heterogeneity.
Econometric methods where identification is based on minimal assumptions provide a ro-
bust structural framework for inference improving credibility and robustness of the empirical
analysis.27 In this context, using pairwise stability as our matchmaking criterion can be used
to develop an empirical framework in the spirit of credible econometrics.28 Since pairwise sta-
bility is the weakest criterion for matchmaking that is consistent with Pareto efficiency, not
specifying the game details allows the econometrician to weaken the behavioral assumptions
that would otherwise be imposed by a specific game.
Our model sheds some light on a puzzle about the recent recovery from the Great Re-
cession in the US. Empirical studies show that wage growth is lower compared to previous
recoveries at the same unemployment rate (Yellen, 2014). Our model predicts that lower fric-
tions imply lower wage growth. If we assume that search frictions in the labor market have
been decreasing over time due to new technologies, our model provides a possible explanation
for this puzzle.
When considering empirical applications, there are a number of ways the model could be
enriched. For example, an application to eBay might consider endogenous search intensity.
An application to the labor market might consider a more realistic production function in-
cluding heterogeneity of both workers and firms, endogenous search intensity and endogenous
entry of firms. The goal of this paper is to construct a parsimonious model that demonstrates
27For example, see Manski (2003), Tamer (2010), Pakes, Porter, Ho, and Ishii (2014).28For example, see Fox (2010a), Fox (2010b), Agarwal (2015), and the references there. See Fox (2009) for
a survey.
30
the importance of network effects in price dispersion, wage growth, allocations of goods and
workers, etc. Enriching the model in other dimensions is an avenue for future research.
ReferencesAgarwal, N. (2015): “An empirical model of the medical match,” American Economic Review, 105(7),
1939–78. (Cited on page 30)
Albrecht, J., P. Gautier, and S. Vroman (2006): “Equilibrium Directed Search with Multiple Appli-cations,” Review of Economic Studies, 73(4), 869–891. (Cited on page 7)
(2014): “Effcient Entry in Competing Auctions,” American Economic Review, forthcoming. (Citedon page 7)
Alvarez, F., and G. Barlevy (2014): “Mandatory Disclosure and Financial Contagion,” Working Paper,Federal Reserve Bank of Chicago. (Cited on page 8)
Ausubel, L. M., and P. R. Milgrom (2002): “Ascending Auctions with Package Bidding,” Frontiers ofTheoretical Economics, 1(1), 1. (Cited on page 7)
Babus, A., and P. Kondor (2013): “Trading and Information Diffusion in Over-the-Counter Markets,”Working Paper. (Cited on page 8)
Bagger, J., F. Fontaine, F. Postel-Vinay, and J.-M. Robin (2014): “Tenure, Experience, HumanCapital, and Wages: A Tractable Equilibrium Search Model of Wage Dynamics,” American EconomicReview, 104(6), 1551–96. (Cited on pages 27, 28, and 43)
Baye, M., J. Morgan, and P. Scholten (2004): “Price Dispersion in the Small and Large: Evidencefrom an Internet Price Comparison Site,” Journal of Industrial Economics, 52, 463–496. (Cited on page 7)
Baye, M. R., and J. Morgan (2001): “Information gatekeepers on the internet and the competitivenessof homogeneous product markets,” American Economic Review, pp. 454–474. (Cited on page 7)
Baye, M. R., J. Morgan, and P. Scholten (2006): “Information, search, and price dispersion,” Handbookon economics and information systems, 1. (Cited on page 7)
Blau, D. M., and P. K. Robins (1990): “Job search outcomes for the employed and unemployed,” Journalof Political Economy, pp. 637–655. (Cited on page 26)
Burdett, K., and K. L. Judd (1983): “Equilibrium price dispersion,” Econometrica, pp. 955–969. (Citedon page 7)
Burdett, K., and D. T. Mortensen (1998): “Wage Differentials, Employer Size, and Unemployment,”International Economic Review, 39, 257–73. (Cited on page 7)
Calvó-Armengol, A., and M. O. Jackson (2004): “The effects of social networks on employment andinequality,” American economic review, pp. 426–454. (Cited on page 5)
(2007): “Networks in labor markets: Wage and employment dynamics and inequality,” Journal ofeconomic theory, 132(1), 27–46. (Cited on page 5)
Carlson, J. A., and R. P. McAfee (1983): “Discrete equilibrium price dispersion,” The Journal ofPolitical Economy, pp. 480–493. (Cited on page 7)
Charness, G., M. Corominas-Bosch, and G. Fréchette (2007): “Bargaining and Network Structure:An Experiment,” Journal of Economic Theory, 136, 28–65. (Cited on page 2)
Corominas-Bosch, M. (2004): “Bargaining in a Network of Buyers and Sellers,” Journal of EconomicTheory, 115, 35–77. (Cited on pages 3, 5, and 6)
Crawford, V. P., and E. M. Knower (1981): “Job Matching with Heterogeneous Firms and Workers,”Econometrica, 49, 437–450. (Cited on pages 7, 19, and 46)
Doval, L. (2014): “A theory of stability in dynamic matching markets,” Working Paper. (Cited on page 19)
31
Einav, L., T. Kuchler, J. Levin, and N. Sundaresan (2015): “Assessing Sale Strategies in OnlineMarkets using Matched Listings,” American Economic Journal: Microeconomics, forthcoming. (Cited onpage 1)
Elliott, M. (2014): “Heterogeneities and the Fragility of Labor Markets,” Mimeo. (Cited on pages 3, 5,and 6)
(2015): “Inefficiencies in networked markets,” American Economic Journal: Microeconomics, forth-coming. (Cited on pages 1, 3, 5, and 6)
Fainmesser, I. P. (2012): “Community structure and market outcomes: A repeated games-in-networksapproach,” American Economic Journal: Microeconomics, 4(1), 32–69. (Cited on page 5)
Fox, J. T. (2009): “Structural empirical work using matching models,” New Palgrave Dictionary of Eco-nomics. Online edition. (Cited on page 30)
(2010a): “Estimating matching games with transfers,” Discussion paper, National Bureau of Eco-nomic Research. (Cited on page 30)
(2010b): “Identification in matching games,” Quantitative Economics, 1(2), 203–254. (Cited onpage 30)
Gale, D., and S. Kariv (2009): “Trading in Networks: A Normal Form Game Experiment,” AmericanEconomic Journal: Microeconomics, 1(2), 114–132. (Cited on page 2)
Gale, D., and L. S. Shapley (1962): “College Admissions and the Stability of Marriage,” AmericanMathematical Monthly, 69, 9–15. (Cited on page 7)
Galenianos, M., and P. Kircher (2009): “Directed search with multiple job applications,” Journal ofEconomic Theory, 144(2), 445–471. (Cited on page 7)
Galeotti, A. (2010): “Talking, searching, and pricing,” International Economic Review, 51(4), 1159–1174.(Cited on page 5)
Gautier, P., P. Muller, B. van der Klaauw, M. Rosholm, and M. Svarer (2014): “EstimatingEquilibrium Effects of Job Search Assistance,” Mimeo. (Cited on page 2)
Gautier, P. A., and C. Holzner (2013): “Simultaneous Search and Network Efficiency,” CESifo WorkingPaper Series 3522, CESifo Group Munich. (Cited on pages 3, 5, and 6)
Gofman, M. (2011): “A Network-Nased Analysis of Over-the-Counter Markets,” Working Paper. (Cited onpage 8)
Hatfield, J. W., and F. Kojima (2010): “Substitutes and Stability for Matching with Contracts,” Journalof Economic Theory, 145, 1704–1723. (Cited on page 7)
Holzer, H. J. (1987): “Job Search by Employed and Unemployed Youth,” Industrial and Labor RelationsReview, 40(4), pp. 601–611. (Cited on page 26)
Jackson, M. O. (2008): Social and Economic Networks, vol. 3. Princeton University Press Princeton. (Citedon page 5)
Jackson, M. O., and A. Wolinsky (1996): “A strategic model of social and economic networks,” Journalof economic theory, 71(1), 44–74. (Cited on page 5)
Janssen, M. C., and J. L. Moraga-González (2004): “Strategic pricing, consumer search and thenumber of firms,” The Review of Economic Studies, 71(4), 1089–1118. (Cited on page 7)
Janssen, M. C., J. L. Moraga-Gonzalez, and M. R. Wildenbeest (2005): “Truly costly sequentialsearch and oligopolistic pricing,” International Journal of Industrial Organization, 23(5), 451–466. (Citedon page 7)
Judd, J. S., and M. Kearns (2008): “Behavioral experiments in networked trade,” in Proceedings of the9th ACM conference on Electronic commerce, pp. 150–159. ACM. (Cited on page 2)
Julien, B., J. Kennes, and I. King (2000): “Bidding for Labor,” Review of Economic Dynamics, 3(4),619–649. (Cited on page 8)
32
Kakade, S. M., M. Kearns, and L. E. Ortiz (2004): “Graphical economics,” in Learning Theory, pp.17–32. Springer. (Cited on page 8)
Kakade, S. M., M. Kearns, L. E. Ortiz, R. Pemantle, and S. Suri (2004): “Economic properties ofsocial networks,” in Advances in Neural Information Processing Systems, pp. 633–640. (Cited on page 8)
Kelso, A. S., and V. P. Crawford (1982): “Job Matching, Coalition Formation, and Gross Substitutes,”Econometrica, 50, 1483–1504. (Cited on pages 7, 19, and 46)
Kircher, P. (2009): “Efficiency of Simultaneous Search,” Journal of Political Economy, 117, 861–913. (Citedon page 7)
Kranton, E., and D. F. Minehart (2001): “A Theory of Buyer-Seller Networks,” American EconomicReview, 91, 485–508. (Cited on pages 3, 5, 6, and 7)
Lester, B. (2011): “Information and prices with capacity constraints,” American Economic Review, 101(4),1591–1600. (Cited on page 7)
MacMinn, R. D. (1980): “Search and market equilibrium,” The Journal of Political Economy, pp. 308–327.(Cited on page 7)
Malamud, S., and M. J. Rostek (2013): “Decentralized Exchange,” University of Geneva. (Cited onpage 8)
Manea, M. (2011): “Bargaining in Stationary Networks,” American Economic Review, 101(5), 2042–80.(Cited on pages 3, 5, and 6)
Manski, C. F. (2003): Partial Identification of Probability Distributions. Springer, New York. (Cited onpage 30)
McAfee, R. P. (1993): “Mechanism Design by Competing Sellers,” Econometrica, 61(6), 1281–1312. (Citedon page 8)
Moen, E. R. (1997): “Competitive Search Equilibrium,” Journal of Political Economy, 105(2), 385–411.(Cited on page 7)
Montgomery, J. D. (1991): “Social networks and labor-market outcomes: Toward an economic analysis,”American Economic Review, pp. 1408–1418. (Cited on page 5)
Mortensen, D. T. (2005): Wage Dispersion: Why Are Similar Workers Paid Differently? The MIT Press,Cambridge. (Cited on page 1)
Morton, F. S., F. Zettelmeyer, and J. Silva-Risso (2001): “Internet car retailing,” The Journal ofIndustrial Economics, 49(4), 501–519. (Cited on page 1)
Myers, C. A., and G. P. Shultz (1951): The dynamics of a labor market: a study of the impact ofemployment changes on labor mobility, job satisfactions, and company and union policies. Prentice-Hall.(Cited on page 5)
Pakes, A., J. Porter, K. Ho, and J. Ishii (2014): “Moment Inequalities and Their Application,” Econo-metrica, forthcoming. (Cited on page 30)
Peters, M., and S. Severinov (1997): “Competition among Sellers Who Offer Auctions Instead of Prices,”Journal of Economic Theory, 75(1), 141–179. (Cited on page 8)
(2006): “Internet auctions with many traders,” Journal of Economic Theory, 130(1), 220–245. (Citedon page 8)
Pissarides, C. A. (1990): Equilibrium unemployment theory. MIT press. (Cited on page 2)
Postel-Vinay, F., and J.-M. Robin (2002): “Equilibrium Wage Dispersion with Worker and EmployerHeterogeneity,” Econometrica, 70, 2295–2350. (Cited on pages 7 and 27)
Rees, A. (1966): “Information networks in labor markets,” American Economic Review, pp. 559–566. (Citedon page 5)
33
Reinganum, J. F. (1979): “A simple model of equilibrium price dispersion,” The Journal of Political Econ-omy, pp. 851–858. (Cited on page 7)
Rogerson, R., R. Shimer, and R. Wright (2005): “Search-Theoretic Models of the Labor Market: ASurvey,” Journal of Economic Literature?, XLIII, 959–988. (Cited on page 6)
Rosenthal, R. W. (1980): “A model in which an increase in the number of sellers leads to a higher price,”Econometrica, pp. 1575–1579. (Cited on page 7)
Roth, A. (2008): “Deferred Acceptance Algorithms: History, Theory, Practice, and Open Questions,” In-ternational Journal of Game Theory, Special Issue in Honor of David Gale on his 85th birthday, 537–569.(Cited on page 7)
Roth, A., and M. Sotomayor (1990): Two-Sided Matching: A Study in Game-Theoretic Modeling andAnalysis. Econometric Society Monograph Series , Cambridge University Press. (Cited on page 7)
Rothschild, M. (1973): “Models of market organization with imperfect information: A survey,” The Journalof Political Economy, pp. 1283–1308. (Cited on page 7)
Salop, S., and J. Stiglitz (1977): “Bargains and ripoffs: A model of monopolistically competitive pricedispersion,” The Review of Economic Studies, pp. 493–510. (Cited on page 7)
Shapley, L., and H. Scarf (1974): “On Cores and Indivisibility,” Journal of Mathematical Economics, 1,11–26. (Cited on page 7)
Shimer, R. (2012): “Reassessing the ins and outs of unemployment,” Review of Economic Dynamics, 15(2),127–148. (Cited on page 26)
Stahl, D. O. (1989): “Oligopolistic pricing with sequential consumer search,” American Economic Review,pp. 700–712. (Cited on page 7)
Stigler, G. J. (1961): “The economics of information,” The journal of political economy, pp. 213–225.(Cited on page 7)
Tamer, E. (2010): “Partial Identification in Econometrics,” Annual Reviews of Economics, 2. (Cited onpage 30)
Varian, H. R. (1980): “A model of sales,” American Economic Review, pp. 651–659. (Cited on page 7)
Walthoff, R. (2012): “Applications and Interviews: A Structural Analysis of Two-Sided SimultaneousSearch,” Working Paper. (Cited on page 7)
Wolinsky, A. (1988): “Dynamic Markets with Competitive Bidding,” Review of Economic Studies, 55(1),71–84. (Cited on page 8)
Yellen, J. L. (2014): “Labor Market Dynamics and Monetary Policy,” Remarks by Chair Janet Yellen atthe Federal Reserve Bank of Kansas City Economic Symposium, Jackson Hole, Wyoming. Available onlineat http://www.federalreserve.gov/newsevents/speech/yellen20140822a.htm [Accessed: 2015 03 19]. (Citedon page 30)
34
Figure 1: Price Distribution: Buyer-Preferred Match.
Notes: Starting in the top left, panels 1 to 4 figure display the empirical distribution of prices from the
model using the buyer-preferred match disaggregated by: i) Market Tightness (which ranges from 0.1 to 5
in the horizontal axis in each graph) and ii) Expected Links per Buyer (1, 2, 3, and 5). Each vertical box
corresponds to a simulated market characterized by those parameters. Each vertical box displays the 95th
percentile (upper whisker), 75th percentile (upper hinge), median (black circle marker), 25th percentile (lower
hinge), and 5th percentile (lower whisker). Note that buyers’ valuation is normalized to range between 0 and
100 which, in turn, bounds the minimum and maximum prices between those values. If the 95th percentile
coincides with the 5th percentile, then the figure shows only a dot (which corresponds to the median too).
In addition, each panel displays the Walrasian outcome, pwalras. We describe how to calculate the Walrasian
outcome in subsection 6.1.
35
Figure 2: Price Distribution: Buyer- vs. Seller-Preferred Matches.
Notes: At each market tightness, panels 1 to 4 display the distribution of prices in the model using the sellers
and the buyer-preferred match. For the seller-preferred match, each vertical box in this figure is identical to
the corresponding vertical box in Figure 1. In addition, each panel displays the Walrasian outcome, pwalras.
We describe how to calculate the Walrasian outcome in subsection 6.1. See the notes in Figure 1 for a
description of the vertical boxes.
36
Figure 3: Price Dispersion and the Walrasian Outcome.
Notes: The top (bottom) figure displays the difference between the 95th (99.5th) price percentile and the
5th (0.5th) price percentile for different market tightness and expected links per buyer using a Nadaraya
Watson kernel regression (of the percentile difference on market tightness) with an Epanechnikov kernel with
bandwidth selected by cross validation. Price distributions are generated using the buyer-preferred match in
a market with 10,000 sellers.
37
Figure 4: The Effect of Frictions on Mean Prices.
Notes: The figure displays the evolution of mean prices over expected links per buyer for different market
tightness. For each market tightness, mean prices are normalized by the mean price when the expected links
per buyer is one. So, by construction, mean prices for each market tightness coincide when the expected links
per buyer is one.
38
Figure 5: Distribution of Matched Buyers.
Notes: Each of the four panels in the figure displays the univariate kernel density estimation of the buyers’
valuations distribution (buyers who bought a good from a seller) for three markets that differ in the ELB
(1, 2, and 5) and for a given market tightness. In addition, each panel displays the distribution of matched
buyers in the Walrasian outcome. We describe how to calculate the Walrasian outcome in subsection 6.1.
Let denote buyers’ valuation in the market. We estimate the probability density function in each market,
f (), as: fK (;h) = 1N h
PN
j=1 K
νν(j)h
, where K (z) is a standard univariate gaussian kernel function,
h is the bandwidth that we choose by cross validation, and (j) , j = 1, . . . , N are the valuations of the
buyers who bought a good in each market. Note that we normalize buyers’ valuation to range between 0
and 100 which, in turn, bounds the minimum and maximum prices between those values. Each valuation
value between 0 and 100, can be interpreted as the percentile for any distribution of buyers’ valuations.
Given that the price distribution has its domain bounded we use a renormalization method to deal with the
boundaries when estimating the productivity probability density function.
39
Figure 6: Wage Distribution in the Labor Market Model (λ = 0.05).
Notes: The figure displays the distribution of wages in the labor market model with = 0.05 (see sub-
section 6.2.1). All figures display the firm-preferred match. Starting in the top left, panel one shows the
empirical distribution of wages for our benchmark, a model with Bertrand competition between at most two
firms (see subsection 6.2.1). Panels 2 to 4 figure display the empirical distribution of wages from the model
using the firm-preferred match disaggregated by: i) Market Tightness (which ranges from 0.1 to 5 in the
horizontal axis in each graph) and ii) Expected Links per firm (1, 2, and 5). Each vertical box corresponds to
a simulated market characterized by those parameters. Each vertical box displays the 95th percentile (upper
whisker), 75th percentile (upper hinge), median (black circle marker), 25th percentile (lower hinge), and 5th
percentile (lower whisker). Note that firms’ productivity is normalized to range between 0 and 100 which,
in turn, bounds the minimum and maximum wages between those values. If the 95th percentile coincides
with the 5th percentile, then the figure shows only a dot (which corresponds to the median too). In addition,
each panel displays the Walrasian outcome, wwalras. We describe how to calculate the Walrasian outcome in
subsection 6.1, where in the case of the labor market wwalras = pwalras.
40
Figure 7: Wage Profile in the Labor Market Model: Mean Wages.
Notes: Each figure displays the wage profile in the labor market model with = 0.05 and the firm-preferred match (see subsection 6.2.1). The horizontal axis
shows the number of periods that the worker has been employed. Mean wages are computed by period. The sample is the set of workers that are employed at least
20 consecutive periods in steady state. Each figure shows the wage profile for a given expected number of links per firm (1, 3, and 5) for different market tightness
(0.5, 1, 3, and 5). As a benchmark, we also display the wage profile for a model with Bertrand competition between at most two firms (see subsection 6.2.1).
41
Figure 8: Mean Wage Growth and Worker Mobility in the Labor Market Model.
Notes: The top figure displays wage growth by market tightness and expected number of links per firm. The
wage growth is defined as the wage of the worker in period 20 minus the wage of the same worker in period
1. We use the sample of workers who have been employed for at least 20 periods. The bottom figure displays
the worker mobility by market tightness and expected number of links per firm. Worker mobility is defined
as the probability that an employed worker makes a job-to-job transition in a period. For both figures, we set
= 0.05 and use the firm-preferred match (see subsection 6.2.1). As a benchmark, we also display the results
for a model restricted to have Bertrand competition between at most two firms (see subsection 6.2.1).
42
Figure 9: Welfare in the Labor Market Model.
Notes: The top panel displays the aggregate production relative to the Walrasian outcome by ELF and by
market tightness. For example, when = 0.5 and ELF=5, the aggregate production is approximately 97.5% of
the aggregate production in the Walrasian outcome. The aggregate production in each market represents the
utilitarian welfare in that market. Figure A5 in the online appendix shows the total aggregate production.
The bottom panel displays the average productivity per worker (total production divided by number of
employees) for different markets. As a benchmark, we also display the results for a model restricted to have
Bertrand competition between at most two firms (see subsection 6.2.1). We set = 0.05 and use the firm-
preferred match (see subsection 6.2.1). Following Bagger, Fontaine, Postel-Vinay, and Robin (2014), we use
the following cumulative distribution function for the productivity of firms F () = 1exp
c1(c0)c2
,
where c0 = 5, c1 = 8, and c2 = 0.7.43
Appendix
A Proof of Theorem 1
In this section we prove Theorem 1.
Theorem 1: Let N = (J , I, E; ν, b) be a network. Let M be a matching. Then the
following are equivalent:
1. There exists pM such that M is stable with respect to pM
2. There exists an abstraction in fully connected graphs A = (G, E; p) such that M is
stable with respect to A.
Proof. 1 implies 2: Let A = (G, E; p) be defined as follows: For each (i, j) 2 M let
Gij = (i, j, (i, j), (j, i)). For each i 2 I such that i(j) = ; let Gi; = (i, ;, ;) and for
each j 2 J such that j(i) = ;, let Gj; = (;, j, ;). Let p(Gij) = pM(i, j), p(Gi;) = b(i) and
p(Gj;) = ν(j). Define E so that (G,G0) 2 E if and only, in the original graph, the buyer in
G is linked with the seller in G0. These ingredients define an abstraction in fully connected
graphs. We need only show that M is stable with respect to A. Since subgraphs are either
trivial or include a single buyer and a single seller, M restricted to G is vacuously stable. Let
G,G0 2 G be such that (G,G0) 2 E. Let j 2 JG and i 2 IG0 . If j(i) = ; then b(i) v(j)
because (M, pM) cannot be blocked. Thus, p(G) = v(j) b(i) = p(G0). If i(j) = ;
then ν(j) v(i) because (M, pM) cannot be blocked. Thus, p(G) = ν(j) v(i) = p(G0).
Otherwise, p(G) = pM(i(j), j), p(G0) = pM(i, j(i)). Then, by stability, p(G) p(G0).
2 implies 1: Let A be as in item 2. Note that, i 2 IG ) v(i) p(G). In words, if i
belongs to a subgraph G, then the virtual price i is receiving must be at least the value of the
subgraph (if i is matched, equality hods, if not, v(i) = b(i) and then b(i) p(G).) Similarly,
j 2 JG ) v(j) p(G). For each edge (i, j) 2 M let pM(i, j) = p(G) where G is such that
(i, j) 2 EG. This is well defined because A does not break M and because M restricted to G
is stable. We must show that no match can be blocked. Let (i, j) 2 M be arbitrarily chosen
and G be such that (i, j) 2 EG. Because M restricted to G is stable, for all (i, j) 2 EG \M ,
v(i) v(j). Now let i0 be such that (i0, j) 2 E but i0 /2 IG; i.e., i0 is linked to j but does
not belong to the same subgraph in the abstraction. Let G0 6= G be such that i0 2 IG0 . Then
(G,G0) 2 E so p(G) p(G0) by cheapest sorting. Thus, v(j) p(G) p(G0) v(i0)
so (i0, j) does not block (M, pM). Since j, i and i0 were arbitrarily chose, then no blocks to
(M, pM) exist, so M is stable with respect to pM .
44
B Formal algorithm and proof of Proposition 1
In this appendix we discuss the formal algorithm used in the main paper and prove
some of its properties. We now present the basic notation we use in the match determination
program. Let st 2 RJI be a matrix of prices for each buyer-seller pair. Each element, sti,j,
represents the price that buyer j would have to bid for seller i if j were to bid for i in round
t. Vector qt represents the bidding queue in period t: qtn = j 2 J represents that in round
t, buyer j is the n-th bidder in the queue. The algorithm ends when l(q) = 0, where l(q)
indicates the length of q. Quantity D(j) indicates j’s demand. Quantities with primes will
indicate quantities that will carry over to the next round of the algorithm. Finally, for each
seller j, we use the following payoff function to model that a buyer can only buy a good from
a seller if the two are linked in the network: uj : I RIJ :! R, uj(i, s) = ν(j) si,j if
(i, j) 2 E and uj(i, s) = 1 otherwise.
Recall some notational conventions: given a matching M , i(·) : J ! I [ ; satisfies
(i(j), j) 2 M for each M -matched j, and i(j) = ; if j is M -unmatched. Analogously,
j(·) : I ! J [ ; satisfies (i, j(i)) 2 M for each M -matched i, and j(i) = ; if i is M -
unmatched. Also, even if not explicitly stated, the network is denoted N = (I,J , E; b, ν(·)),
I = #I, J = #J .
Match Determination Program.
Input= (N , s0, (u1, ..., uJ), h0, q) where:
• s0 = (s01, ..., s0J) 2 R
JI , s0j = (b, ..., b) 2 RI ,
• For each buyer j, and each t 2 N [ 0, uj(i, st) = ν(j) sti,j if (i, j) 2 E and
uj(i, st) = 1 if (i, j) /2 E,
• h0 = (0, ..., 0) 2 RIJ .
• q0 2 J J such that q0n = q0m iff m = n.
Start step R(1):
R(t). Set ht = h, st = s, qt = q, j = q1.
1. If maxuj(i, s) : i 2 I < 0 set s0 = s and h0 = h, q0 = (q2, ..., ql(q)).
a. If l(q0) = 0, stop, set M = (i, j) : hi,j = 1, and Output= M .
b. If l(q0) 6= 0, set qt+1 = q0, st+1 = s0, ht+1 = h0 and proceed to R(t+ 1).
2. If maxuj(i, s) : i 2 I 0 let D(j) 2 argmaxuj(i, s) : i 2 I.
a. If argmaxuj(i, s) : i 2 I has more than one element, select D(j) 2 argmaxi2Iuj(i, s)
randomly.
45
3. Set the following parameters:
a. s0D(j),j = sD(j),j; for all j0 6= j, s0D(j),j0 = sD(j),j +∆
2; s0i00,j00 = si00,j00 elsewhere,
b. If hD(j),j0 = 0 for all j0 6= j, set q0 = (q2, ..., ql(q)); if hD(j),j0 = 1 for some j0 6= j, set
q0 = (q2, ..., ql(q), j0),
c. h0D(j),j = 1; for all j0 6= j, h0
D(j),j0 = 0; h0i00,j00 = hi00,j00 elsewhere.
4. If l(q0) = 0, stop. Set M = (i, j) : hi,j = 1. Output= M .
If l(q) 6= 0 set h0 = ht+1 and s0 = st+1 and q0 = qt+1. Then start R(t+ 1).
Although this algorithm is motivated by Crawford and Knower (1981) and Kelso and Craw-
ford (1982), there are three important differences. The first is that firm productivities increase
in increments of ∆ whereas bids increase in increments of ∆
2. Since Crawford and Knower
(1981) and Kelso and Crawford (1982) work with a discrete core, the algorithm they run
produces a stable match when both bid increments and productivities increase by the same
amount. However, since we work with a continuous core, it is not true that the matching
generated by such an algorithm is stable. One can construct examples where the matching
generated by the algorithms in Crawford and Knower (1981) and Kelso and Crawford (1982)
(say, M) satisfies that there is no price function pM such that M is stable with respect to
pM . We provide one example in section B (Figure A6) in the online appendix. The modifi-
cation we introduce, that bids live in a finer grid than firm productivities, helps us bypass
this problem. The second difference with the algorithms in Crawford and Knower (1981)
and Kelso and Crawford (1982) is that we only use their program to find the matching, but
not the prices that make it stable. The reason is that their algorithm makes prices rise too
quickly. While in some networks the price generated by the algorithms in Crawford and
Knower (1981) and Kelso and Crawford (1982) is the pointwise minimum price that makes
the matching stable, this is not always guaranteed. This is because in our setting we violate
the non-indifference assumptions made in Crawford and Knower (1981) and Kelso and Craw-
ford (1982). In order to capture, for each matching, the pointwise maximum and minimum
prices at which that matching is stable we run two independent programs. We call these the
Price Determination Programs, and we describe them below. The first Price Determination
Program (I), outputs the pointwise minimum price function at which a matching is stable.
The second Price Determination program (II), outputs the pointwise maximum price func-
tion at which a matching is stable. The third difference is that, when a seller i accepts a bid
from a buyer j, then any future bid buyer j0 submits to i must outbid j’s bid. In symbols, if
in round t seller i accepts bid sti,j from j, then at the end of round t all sellers j0 linked to i
have their bid price raised to st+1i,j0 = sti,j +
∆
2. This modification reduces the run time of the
algorithm by a factor of four.
Price Determination Program (I).
Input= (N ,M).
46
1. For each i 2 I such that j(i) = ; set ρ1i = b.
2. For each i 2 I such that j(i) 6= ; set ρ1i = maxν(j) : (i, j) 2 E and i(j) = ;.
3. Set t = 1. Start step 4(1).
4(t). Given (ρt1, ..., ρtI):
a. For each i 2 I such that j(i) = ; set ρt+1i = ρti.
b. For each i 2 I such that j(i) 6= ;, let j j(i). Then, set
ρt+1i = maxρti0 : (9j
0)(i0, j0) 2 M , (i, j0) 2 E.
c. If for all i 2 I ρt+1i = ρti:
For each i such that j(i) 6= ; set pM(i, j(i)) = ρt+1i .
Output= (pM(·)).
d. Otherwise, start step 4(t+ 1).
Price Determination program (I) outputs the minimum price at which M can be made
stable.
Price Determination Program (II).
Input= (N ,M).
1. For each i 2 I such that j(i) = ; set ρ1i = b.
2. For each i 2 I such that j(i) 6= ; set ρ1i = ν(j(i)).
3. Set t = 1. Start step 4(1).
4(t). Given (ρt1, ..., ρtI):
a. For each i 2 I such that j(i) = ; set ρt+1i = ρti.
b. For each i 2 I such that j(i) 6= ;, let j j(i). Then, set
ρt+1i = minρti0 : (i
0, j) 2 E.
c. If for all i 2 I ρt+1i = ρti:
For each i such that j(i) 6= ; set pM(i, j(i)) = ρt+1i .
Output= (pM(·)).
d. Otherwise, start step 4(t+ 1).
47
Price Determination program (II) outputs the maximum price at which M can be made
stable.
In Section 5 we claimed our algorithm has four properties: it ends in finite time, it
selects a pairwise stable allocation, and for each allocation it selects the pointwise minimum
and maximum prices that sustain it.
Proposition 1: The deferred acceptance algorithm has the following properties:
1. It stops after a finite number of rounds.
2. It outputs a pairwise stable allocation.
3. Price Determination program (I) outputs the pointwise minimum price function at
which M is stable.
4. Price Determination program (II) outputs the pointwise maximum price function at
which M is stable.
We now prove these items one at a time. In what follows, we use MDP and PDP
to abbreviate the Matching Determination Program and the Price Determination Program
respectively. Finally, if (xi)i2I is a vector indexed by I we use the convenient shorthand
notation x· to denote the whole vector, whenever ambiguity is unlikely.
We need two lemmas: the first, shows that, given M produced by the MDP, there
exist prices pM such that M is stable with respect to M . The second shows that the prices
generated by the PDP are weakly lower than any pM such that M is stable with respect to
M . To prove these Lemmas, recall that (ρti)i2I,t1 from the PDP(I) is defined as follows:
• If j(i) = ;, ρti = b for all t.
• If j(i) = j for some j 2 J , ρ1i = maxν(j) : (i, j) 2 E, i(j) = ; for each i 2 I, and
ρti = maxρt1i0 : (9j0, i0) : (j0, i0) 2 M , (j0, i) 2 E for all t 2.
The following properties imply that there exists a value T such that, for all i and all t T ,
ρti = ρt+1i . That is, (ρt·)t0 is eventually constant. We let ρ1· limt!1 ρt· .
1. For all i, ρti ρt+1i . This follows because ρt1
i 2 ρt1i0 : (9j0) : (j0, i0) 2 M , (j0, i) 2 E
whenever j(i) = j and ρti = b whenever j(i) = ;.
2. For all i, ρti maxν(j) : j 2 J .
3. For all i, if ρti 6= ρt+1i then ρt+1
i ρti ∆.
48
Finally, recall that ∆ 2 R is chosen so that for all j 2 J , ν(j) = b+kj∆ for for kj 2 N[0.
In particular, ν(j) b for all j. This normalization only rules out uninteresting cases where
a buyer never places a bid and is never matched to a seller.
Lemma 1. Let M be the matching produced by the MDP. Then, there exists pM such that
M is stable with respect to pM .
Proof. For each edge (i, j) 2 M define pM(i, j) = ρ1i where (ρti)i2J ,t2N[1 is as defined in
the PDP(I). Also, let T be the last round of the MDP and let [sTi,j]i2I,j2J be the matrix of
final prices generated by the MDP. We show that M is stable with respect to pM . Assume
first that (i, j) 2 E are such that j(i) = i(j) = ;. Then i received no bids, so sTi,j = b.
Since the algorithm ended, it must be that uj(i, sT ) < 0 , ν(j) < b, a contradiction. Thus,
there does not exist an edge (i, j) 2 E such that j(i) = i(j) = ; so, a fortiori, no such
edge (i, j) 2 E blocks M . Now let (i, j) 2 M . Pick j0 6= j such that (i, j0) 2 E. We show
(i, j0) 2 E does not block M . If i(j0) = ; then ν(j0) ρ1i ρ1i = pM(i, j). If i(j0) 6= ;
then ρ1i ρ1i(j0) by construction. Thus, pM(i, j) pM(i(j0), j0). Thus, (i, j0) does not block
M . Pick i0 6= i such that (i0, j) 2 E. We show (i0, j) 2 E does not block M . If j(i0) 6= ;
then ρ1i0 ρ1i by construction. Thus, pM(i0, j(i0)) pM(i, j). Let j(i0) = ;. Then i0 never
received a bid. Let t be the last time j bids for i. Since bidders bid for the cheapest seller
sti,j sti0,j = b. By definition of t, sti,j = sTi,j so sTi,j = pM(i, j) = b. We use this to argue that
ρ1i= b for all matched i such that (i, j(i)) 2 E (note that i is one such i). Pick i such that
j(i) 6= ; and (i, j(i)) 2 E. Then, sTi,j (i)
sTi,j (i)
. 29 Since sTi,j = b and sTi,j (i)
sTi,j +∆
2,
then sTi,j (i)
b+ ∆
2. If there exists j such that i(j) = ; and (i, j) 2 E, then it must be that
ν(j) = b. Indeed, if ν(j) > b then ν(j) b+∆ which is a contradiction: since sTi,j (i)
b+ ∆
2
and i(j) = ;, uj (i, sT ) 0, which contradicts T being the last round of the MDP. Thus,
ν(j) = b. Hence, ρ1i= b. We now conclude the argument in an inductive manner: if ρk
i= b
for some k and all i that satisfy j(i) 6= ; and (i, j(i)) 2 E, then by construction ρk+1i = b.
Thus, ρ1i = b = pM(i, j). Thus, (i0, j) does not block M . Therefore, M is stable with respect
to pM .
Lemma 2. Let M be the matching generated by the MDP. Let pM be any price function such
that M is stable with respect to pM (which is well defined by our previous lemma) and let v
be the associated payment function. Let pM be the price generated by the PDP(I) and v the
corresponding payment function. Then, v v.
Proof. Let M , pM , v, pM and v be as in the statement of the lemma. Then, for all i,
v(i) ρ1i . Indeed, if j(i) = ; then v(i) = b = ρ1i . If j(i) = j for some j then, by stability
of M with respect to pM , v(i) ν(j0) for each j0 such that i(j0) = ;. Thus, v(i) ρ1i .
29 Indeed, let t be the last time j(i) bids for i. Then, sti,j∗ (i)
= sTi,j∗ (i)
, and sti,j∗ (i)
sti,j∗ (i)
, where the
last inequality holds because buyers always bid for the cheapest sellers. By monotonicity of the matrix ofprices, st
i,j∗ (i) sT
i,j∗ (i). Thus, sT
i,j∗ (i) sT
i,j∗ (i).
49
We now show that if v ρt for some k, then v ρt+1. Indeed, for all i such that j(i) = ;,
v(i) = b = ρti = ρt+1i . For all i such that j(i) = j, we have the following:
ρt+1i = maxρti0 : (9j
0, i0)(i0, j0) 2 M and (i, j0) 2 E
maxv(i0) : (9j0, i0)(i0, j0) 2 M and (i, j0) 2 E v(i),
where the last inequality follows from stability of M with respect to pM . Thus, for each t
and each i, ρti v(i). Hence ρ1· v(·) v(·).
We now prove items 1 through 4 of Proposition 1.
1. The algorithm ends in finite time.
Proof. By the same arguments as Crawford-Knoer, the matching determination program
ends in finite time. Furthermore, let K 2 N satisfy maxν(j) : j 2 J = b+K∆. Then the
price determination program ends in at most 2K rounds.
2. The algorithm outputs a pairwise stable matching.
Lemma 1 already shows that M is stable with respect to pM when pM is the price function
generated by PDP(I). Lemma 3 below shows that M is also stable with respect to pM when
pM is the price function generated by PDP(II).
3. Price Determination program (I) outputs the pointwise minimum price
function at which M is stable.
Proof. Follows from lemma 2 and that M is stable with respect to pM , where pM is the price
function generated by the PDP (I).
4. Price Determination program (II) outputs the pointwise maximum price
function at which M is stable.
The proof is analogous to the proof that the PDP(I) outputs the pointwise minimum price
function at which M is stable. Reasoning as in Lemma 2, if pM is such that M is stable at
prices pM , and v is the corresponding payment function, then v(·) ρ1· (Lemma 4 below).
Moreover, M is stable at prices induced by ρ1· (Lemma 3, below). The result then follows
from Lemmas 3 and 4.
Lemma 3. Let M be the matching generated by the MDP, and let pM be the prices generated
by the PDP(II). M is stable with respect to pM .
Proof. Let M be the matching outputted by the matching determination program, and pM
be the prices generated by the price determination program. Assume (i, j) 2 E, j(i) =
i(j) = ;. Since there exists pM such that M is stable with respect to pM then ν(j) b.
50
Thus (i, j) do not block M at ρ1. Now consider (i, j) 2 M . We show no seller and no buyer
wishes to block (i, j):
a. No Buyer blocks: Let j0 be such that (i, j0) 2 E. If i(j0) 6= ; then, by construction,
ρ1i(j0) ρ1i , so (i, j0) does not block. Assume now that i(j0) = ;. We say a seller i0
is indirectly connected to seller j if there exists sequences (i1, ..., ik) and (j1, ..., jk1)
such that (i1, j) 2 E, (i1, j1) 2 E, (i2, j1) 2 E, ..., (ik, jk1) 2 E, with i0 = ik.
That is, if a path can be constructed from j to i0. By construction, minν(j(i0)) :
i0 is indirectly connected to j ρ1i where, by convention, ν(;) = b. Now consider
the abstraction used in Theorem 1 item [1]: each matched pair (i, j) 2 M is assigned
their own subgraph, and all unmatched buyers/sellers are assigned a trivial subgraph
that contains only them. Because there exist prices pM such that M is stable at pM ,
cheapest sorting implies that
ν(j0) pM(i, j) minv(i0) : i0 is indirectly connected to j
minν(j(i0)) : i0 is indirectly connected to j.
Thus, ν(j0) ρ1i so (i, j0) does not block.
b No Seller blocks: Let i0 be such that (i0, j) 2 E. By construction, ρ1i ρ1i0 . Thus,
(i0, j) does not block.
Lemma 4. Let M be the matching generated by the MDP. Let pM be any price function such
that M is stable with respect to pM (which is well defined by our previous lemma) and let v
be the associated payment function. Let pM be the price generated by the PDP(II) and v the
corresponding payment function. Then, v v.
Proof. Let M , pM , v, pM and v be as in the statement of the lemma. Then, v(i) ρ1i for
all i. Indeed, if j(i) = ; then v(i) = b = ρ1i . If j(i) = j for some j then, by stability of M
with respect to pM , v(i) ν(j) = ρ1i .
We now show that if v ρt for some k, then v ρt+1. Indeed, for all i such that j(i) = ;,
v(i) = b = ρti = ρt+1i . For all i such that j(i) = j, we have the following:
ρt+1i = minρti0 : (i
0, j) 2 E
minv(i0) : (i0, j) 2 E v(i),
where the last inequality follows from stability of M with respect to pM . Thus, for each t
and eaxh i, ρti v(i). Hence ρ1· v(·) v(·).
51